OpenRouter
面向 AI Agent 栈的中立多模型路由层
OpenRouter 在多模型 AI 技术栈中确有战略位置,但公开证据仍不足以高置信度承销据报 $1.3B 的估值。
封面要素
公司概况
OpenRouter 是一家私营 AI 基础设施公司,卖的是统一 API、路由层和治理界面,覆盖数百个大语言模型与多模态模型。产品为开发者、agent 构建者和企业 AI 团队统一访问、定价、故障转移和供应商选择;这些客户不想把系统写死在单一模型厂商上。
- 成立时间
- 2023-01-01
- 创始人
- Alex Atallah, Louis Vichy
- 创立地点
- New York, NY, USA
- 总部
- New York, NY, USA
- 产品
- 兼容 OpenAI 的多模型 API,覆盖路由、故障转移、供应商选择、模型排名、分析、护栏、隐私控制、SDK,以及不断扩展的 agent / 服务器工具能力。
- 客户
- 开发者、AI 初创公司、编程 agent 构建者,以及部署多模型 AI 应用的企业团队
- 商业模式
- 围绕预付积分、BYOK 费用和企业合同搭建的用量计费路由与平台费模式
- 阶段
- Series B / Growth
- 融资情况
- 2026 年 5 月由 CapitalG 领投的 $113M B 轮,报道估值约 ~$1.3B;公开证据支持至少 $153M 已披露股权融资,但历史总融资仍披露不足
执行摘要
主要优势
- 中立的多模型 API 和路由位置,覆盖 400+ 个模型,一份合同、一张账单,并保持广泛开发者兼容性
- 截至 2026 年 5 月,每周 25T token 调用量和 8M+ 用户规模,释放出很强的公开需求信号
- CapitalG 和重磅风投财团参与 2026 年 Series B,提供了战略验证
主要风险
- 尽管估值已到独角兽水平,当前收入、毛利率、净留存和员工数仍未披露
- 原生云路由、模型提供商和开源网关可能压缩抽成率,并绕过路由层
- 可靠性、隐私或提供商日志记录一旦出问题,会直接削弱 OpenRouter 的核心信任主张
未决问题
- 当前 ARR 或收入运行率未公开披露,估值支撑受限
- 融资轮重叠、清算优先权、老股占比和累计融资总额仍不清楚
- 客户集中度、续约行为、毛利率和员工数在公开材料中仍不透明
目录
01公司概览
1.1 身份定位、产品与运营模式
OpenRouter 把自己定位成基础设施,而不是消费应用或模型实验室。从官网、快速入门 文档、投资人文案到企业营销材料,公司都反复使用同一套说法:它是 LLM 的统一接口,也是 AI 模型交易所;一个 API、一个计费层、一套路由网络,横跨数百个模型和数十家供应商。这个定位很关键,因为它同时解释客户承诺和商业模式。客户买的是摆脱直接厂商锁定的抽象层,也买自动故障转移、路由优化、治理控制和用量报告。企业层现在已经超过基础路由。Workspaces、零数据保留控制、支出上限、供应商白名单、护栏和可观测性都表明,OpenRouter 想做生产 AI 流量的控制平面,而不是给开发者用的便利封装。法律和运营足迹比组织架构更清楚:公开条款和镜像 Form D 指向 OpenRouter, Inc.,一家带纽约地址的 Delaware 公司;招聘页则把团队描述成远程优先、在美国各地招聘的小团队。这个组合说明,公司仍然精简,却在远超公开员工规模所暗示的用量层级上出售基础设施。[CO001, CO002, CO003, CO004, CO005, CO006]
| 指标 | 数值 / 状态 | 日期 | 信心 | 缺口 / 注意点 |
|---|---|---|---|---|
| 法律实体 / 创立年份 | OpenRouter, Inc. / 2023 | 2023 | 高 | 2023 年创立有充分依据;具体月份未在留存来源中公开。 |
| 主要营业地址 | 注册地址:169 Madison Avenue, New York, NY 10016 | 2025-2026 | 高 | 地址依据来自法律页面和备案镜像,而不是公司关于页面页脚。 |
| 当前阶段 | 私有成长阶段公司;Series B 后 | 2026-05 | 高 | 阶段由 2026 年 5 月 Series B 及未发现公开上市或 S-1 证据推断。 |
| 最新融资 | $113M Series B,由 CapitalG 领投 | 2026-05-26 | 高 | 轮次规模清楚;确切交割机制和是否有老股转让仍不清楚。 |
| 最新估值 | 约 $1.3B 投后估值 | 2026-05-26 | 高 | 估值来自媒体报道,而不是公司备案。 |
| 可支撑的公开披露融资额 | 公开轮次报道合计至少 $153M | 2025-2026 | 中 | Form D 可能与所报道 Series A 重叠,无法精确计算历次融资总额。 |
| 收入 / 年化收入 | 2026-07-01 | 低 | 留存来源中未找到公开收入、ARR、毛利率或 NRR 披露。 | |
| 使用规模 | 每周 25T token / 每月 100T token / 8M+ 用户 | 2026-05 | 高 | 这些是公司口径,并非经审计的运营指标。 |
| 当前官网计数 | 10M+ 用户 / 70+ 供应商 / 400+ 模型 | 2026-07-01 | 中 | 首页计数是当前营销指标,与 2026 年 5 月用户数略有冲突。 |
| 员工数 | 2026-07-01 | 低 | 招聘页面称团队小且远程优先,但没有披露精确员工数。 | |
| 债务 / 授信安排 | 2026-07-01 | 低 | 未留存到创业债、仓储融资额度或其他授信安排的公开证据。 |
数值保留留存公开记录实际能支撑的内容。空值表示截至运行日,备案、官方页面或独立报道无法支撑该指标。
[CO001, CO002, CO006, CO014, CO016, CO018]身份、产品、控制、生态、资本和风险,都指向同一个多模型基础设施论点。
[CO003, CO004, CO005, CO025, CO026, CO027]1.2 创始人、管理层与治理
管理层能见度集中在 Alex Atallah 身上。BusinessWire、TechCrunch 相邻报道和申报镜像都支持他是 CEO 兼联合创始人,申报文件也显示他担任董事。更难的是创始人名单是否完整。The SaaS News 把 Louis Vichy 列为联合创始人,但这条归因比 Alex Atallah 的记录薄,还没有强官方创始人页或广泛高管名单支撑。公开治理能见度同样窄。镜像 Form D 列出 Anjney Midha 为董事,但保留的公开来源没有给出完整董事会名单、委员会结构、观察员权利,或当前投资人附带的控制条款。因此,本章只能描述治理锚点,不能精确量化创始人控制。关键人依赖的故事仍然直观:Atallah 看起来是产品愿景、融资和市场叙事的核心公开代言人;更宽公开高管梯队缺失,则带来继任与扩张尽调问题,管理层需要直接回答。[CO008, CO009, CO010, CO011, CO012, CO013]
| 人物 | 角色 | 背景 / 语境 | 职能覆盖 | 关键人依赖 |
|---|---|---|---|---|
| Alex Atallah | CEO、联合创始人、董事 | 公开承载融资、市场叙事和经营策略;镜像 Form D 也将其列为高管和董事。 | 公司战略、融资、产品叙事、最高层治理 | 高 |
| Louis Vichy | 联合创始人(二手来源归属) | 创业媒体将其列为联合创始人,但当前公开佐证比 Atallah 薄。 | 可能覆盖产品 / 技术创始职能,但公开职责披露不足 | 中 |
| Anjney Midha | 董事 | 镜像 Form D 将其列为董事,是管理层之外一个公开可见的董事会级治理锚点。 | 董事会监督与投资人治理连接 | 低 |
该表只纳入留存公开证据可支撑的创始人和董事姓名,不是完整领导层组织图。
[CO008, CO009, CO010, CO011, CO012, CO013]1.3 融资历史、资本结构与规模信号
融资记录在最新一轮最强,在历史总额上最弱。OpenRouter 于 2026 年 5 月下旬宣布 CapitalG 领投的 $113M B 轮,TechCrunch 和其他创业媒体把该轮与约 $1.3B 的投后 估值相连。财团把传统风投与战略平台投资人混在一起,包括 NVentures、ServiceNow Ventures、MongoDB Ventures、Snowflake Ventures 和 Databricks Ventures,这强化了 OpenRouter 想嵌入企业 AI 基础设施栈的信息。TechCrunch 还报道,2025 年 6 月公司完成 Andreessen Horowitz 和 Menlo Ventures 领投、Sequoia 参投的 $40M A 轮。这支持至少 $153M 的公开报道股权融资,但精确历史总额仍无法干净对齐,因为一份 2025 年 8 月镜像 Form D 显示,发行规模更大,为 $47.6M,已售 $29.6M,投资人 30 名。没有底层融资文件,投资人无法判断该申报与后来媒体概述的轮次重叠多少。规模信号更清晰:官方材料多次引用每周 25T tokens、每月 100T tokens、超过 8M 用户、超过 400 个模型,以及后续网站计数器显示超过 10M 用户、超过 70 家供应商。这些是公司说法,不是审计 KPI,但足以证明 OpenRouter 已在极大用量下销售,尽管收入和员工数仍未披露。[CO014, CO015, CO016, CO017, CO018, CO019]
| 利益相关方 | 角色 | 控制权或经济重要性 | 尽调问题 |
|---|---|---|---|
| CapitalG | Series B 领投方 | 锚定 2026 年轮次,并可能对成长阶段治理和商业信号有更大影响。 | 要求披露董事席位、按比例跟投权和信息权细节。 |
| Andreessen Horowitz | 现有投资人 | 仍在 Series B 公告中列名,此前也与 2025 年 Series A 相关。 | 确认 Series B 后持股,以及任何治理保护。 |
| Menlo Ventures | 现有投资人 | 在此前轮次报道中与 a16z 并列,也在 Series B 公告中作为现有支持方列名。 | 确认 Menlo 是否仍有董事会代表。 |
| NVentures | Series B 战略参与方 | NVIDIA 关联投资人增强了 OpenRouter 的基础设施和 AI 平台可信度。 | 澄清任何商业绑定、优先供应商经济条款或信号权。 |
| ServiceNow Ventures | Series B 战略参与方 | 如果商业关系出现,可能成为分发或企业可信度桥梁。 | 询问是否存在产品集成或市场拓展权利。 |
| MongoDB Ventures | Series B 战略参与方 | 显示对应用栈基础设施买方的相关性。 | 询问该投资是纯财务投资,还是包含生态承诺。 |
| Snowflake Ventures | Series B 战略参与方 | 为企业 AI 工作负载增加数据平台邻近性。 | 澄清是否有联合路线图或数据治理合作。 |
| Databricks Ventures | Series B 战略参与方 | 对企业 AI 工作流分发和生态对齐可能有价值。 | 确认任何技术或市场拓展合作。 |
| Anjney Midha | Form D 中列名的董事 | 管理层之外一个公开可见的治理角色。 | 索取完整董事会名单、观察员名单和委员会图谱。 |
该图谱强调对控制权和商业杠杆最重要的公开列名投资人和治理角色,不是完整股权结构表。
[CO012, CO014, CO015, CO017, CO019, CO021]公开记分牌主要由融资和使用量指标主导,而员工数和收入仍不透明。
用户数和模型供应商数反映公司在不同发布时点披露的运营指标。收入和员工数保持定性,因为公开记录无法支撑精确数字。
[CO016, CO018, CO022, CO023, CO024, CO046]1.4 里程碑、风险信号与尽调缺口
OpenRouter 2026 年的里程碑节奏显示,公司正从路由工具快速走向更宽的基础设施平台。Workspaces 增加组织级控制,5 月发布重点把护栏和 PII 脱敏推入核心产品,图像 API 扩展到媒体生成,6 月 MCP 发布又把 OpenRouter 直接带入编程 agent 工作流。这些发布增强了运营论点,因为产品从 API 访问延伸到治理、工具和生态分发。但同一时期也暴露了最清楚的公开负面信号。OpenRouter 详细复盘了 2026 年 2 月两次与第三方缓存依赖有关的故障,包括客户侧 500 和误导性 401。透明度是加分项,但事故仍削弱公司最核心的可靠性承诺。独立评论又加一层警示:TrueFoundry 对用户投诉的综合指向支持响应、账户安全处理,以及自主 agent 工作负载下偏弱的硬预算控制。最后,第三方目录数据噪音很大,值得主动怀疑。一份归档 Tracxn 页面仍把公司描述为 Anguilla 基地、且仅由 Soma Capital 支持,这与后来的纽约申报证据和更广的 2026 年投资人阵容直接冲突。总体看,公开故事可投、增长很快,但披露还不完整。[CO029, CO030, CO031, CO032, CO033, CO034]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2023 | OpenRouter 创立 / 2023 年初启动 | 创立 | 公司成立 | OpenRouter, Inc.;Alex Atallah;后续创业媒体的创始人归属包括 Louis Vichy | 为极短的扩张时间线设定基准。 |
| 2025-05-27 | 镜像 Form D 发行的首次销售日期 | 融资 | $47.6M 发行启动 | OpenRouter;外部投资人 | 显示募资活动在 8 月备案日期前已经开始。 |
| 2025-08-25 | Form D 提交 | 监管 | 向 30 名投资人售出 $29.6M | OpenRouter;SEC 备案镜像;Alexander Atallah 签署 | 补上主要营业地址和治理线索,但总融资额仍不清楚。 |
| 2026-02-17 | 首次缓存相关宕机 | 不利事件 | 38 分钟严重中断 | OpenRouter;第三方缓存供应商;客户 | 考验公司最核心的可靠性主张。 |
| 2026-02-19 | 第二次相关宕机并确认根因 | 不利事件 | 35 分钟严重中断 | OpenRouter;第三方缓存供应商 | 倒逼围绕熔断器和更清晰错误码修复。 |
| 2026-Q2 | Workspaces 上线并发布文档 | 治理 | 企业控制面扩展 | OpenRouter 组织与管理员 | 显示定位从开发者便利转向企业管理。 |
| 2026-05-26 | CapitalG 领投的 Series B 宣布 | 融资 | $113M,投后估值约 $1.3B | CapitalG;NVentures;ServiceNow Ventures;MongoDB Ventures;Snowflake Ventures;Databricks Ventures;现有投资人 | 确立 OpenRouter 作为独角兽级 AI 基础设施公司的定位。 |
| 2026-05 | 5 月发布重点和 100T 月度 token 量 | 规模 | 每月 100T token | OpenRouter | 把融资与可见经营规模和治理功能扩展连接起来。 |
| 2026-06 | Unified Image API 发布 | 合作 | 30+ 图像模型;按端点定价 | OpenRouter 及包括 Google、OpenAI、Microsoft、xAI 等在内的模型供应商 | 显示供应商网络从文本推理扩展到多模态媒体工作流。 |
| 2026-06 | Releasebot 突出 MCP 服务器发布 | 产品 | 实时模型数据和编码代理工具 | OpenRouter;编码代理生态 | 加深在开发者工作流和模型选择工具中的分发。 |
该时间线捕捉了会实质影响尽调的主要公开里程碑。它不是每一次发布、客户或治理事件的完整内部时间线。
[CO007, CO014, CO019, CO029, CO030, CO031]从 2023 年成立,到 2026 年产品扩展、融资和可靠性压力的压缩时间线。
[CO007, CO014, CO019, CO029, CO030, CO031]1.5 证据要点
02市场分析
2.1 市场边界、计入支出与现状替代方案
OpenRouter 并不覆盖整个生成式 AI 软件栈。最贴近它的市场,是位于应用与外部模型供应商之间的托管多模型路由层:统一 API、选择供应商、设置回退,并越来越多地加入缓存、分析和计费控制。这个边界比简单模型选择器更宽,但比完整企业 AI 控制平面更窄。计入支出因此包括第三方模型上的统一 API 访问、供应商路由、故障转移、缓存、用量分析和合并账单。排除支出包括训练和模型创建、GPU 硬件、与 AI 推理无关的通用企业 API 项目,以及终端应用 AI 收入。邻近集合很重要,因为 OpenRouter 对抗的是几种现状:开发者可以直接调用供应商;平台团队可以扩展 Cloudflare、Kong 或 Azure API Management;云原生团队可以留在 AWS Bedrock 或 Microsoft Foundry;自托管团队可以运行 LiteLLM。因此,OpenRouter 真正可服务的市场,是那些看重快速多供应商接入、但不想承担自托管或深度基础设施所有权的那一片。[CM001, CM002, CM003, CM004, CM015, CM016]
| 细分 / 类别 | 纳入支出 | 排除支出 | 主要买方 / 付款方 | OpenRouter 相关性 |
|---|---|---|---|---|
| 托管式多模型路由 SaaS | 第三方模型的统一 API 接入、模型选择、故障回退、缓存、用量分析和统一计费 | 模型训练、基础模型研发或专用 GPU 基础设施 | 先由开发者买单;后续进入 AI 平台或平台工程预算 | 核心可寻址层 |
| 企业 AI 网关 / 控制平面 | 治理、RBAC、配额、审计日志、防护栏、token 政策、MCP 或代理控制 | 缺少组织控制的纯模型实验 | 平台工程、中央 AI 平台、安全、IT 治理 | 相邻层,在企业账户中常常重叠 |
| 云原生托管路由器 | Bedrock 提示词路由、Foundry 模型路由器、云 AI 服务治理 | 跨云独立性或供应商无关计费 | 云平台负责人和云原生应用团队 | 主要替代品,限制第三方渗透 |
| AI 化 API 管理 | 统一模型端点、策略执行、token 配额、日志、负载均衡、熔断 | 模型训练、前沿模型创建或终端用户 AI 应用 | API 平台、集成或安全团队 | 当 AI 位于既有 API 体系内时,是强替代品 |
| 自托管开源路由 | 自有基础设施上的供应商抽象、路由逻辑、重试、负载均衡、回退和成本控制 | 托管 SaaS 便利性和外包运营 | 平台工程或开发者基础设施团队 | 对受监管或成本敏感团队是重要替代品 |
| 排除的相邻层 | 终端应用 AI 软件、GPU 硬件、推理芯片、训练集群和非 AI API 项目 | n/a | n/a | 不在 OpenRouter 特定 SAM 内 |
边界包括围绕外部模型使用的路由、故障切换、缓存、可观测性和计费控制支出,但排除模型创建、硬件和大多数非 AI API 管理支出。
[CM001, CM002, CM003, CM015, CM016, CM017]2.2 TAM/SAM/SOM 口径与保留矛盾
OpenRouter 所处类别的公开市场测算高度依赖定义。最窄口径是 Intel Market Research 的 LLM middleware gateway 市场,隐含 2026 年多模型中间件收入仅 $18.9M。更宽口径上升很快:Intel 的 AI API gateway 市场 2026 年达到 $0.85B;The Business Research Company 估算 LLM gateways 2026 年为 $2.76B,更宽的 LLM gateway platform 市场 2026 年为 $4.23B;Research and Markets 则给出更宽 AI gateway 口径,2025 年已达 $3.66B。这些估计不应被平均抹平。它们测量的边界不同,而且一个 TBRC 页面甚至保留了内部不一致:同一个 $11.01B 预测,在概览中称为 2030 年结果,在报告属性框中又称为 2035 年预测。实务结论是,OpenRouter 的真实可服务市场不是整个 gateway 或 AI middleware 宇宙。更受证据约束的 2026 年 SAM 是托管多模型路由 SaaS 的低数亿美元切片;扣除原生云、企业网关和自托管栈已经拿走的需求后,近期 SOM 更小。[CM007, CM008, CM009, CM010, CM011, CM012]
| 发布方 | 年份 | 地理范围 | 价值 / 范围(USD B) | CAGR | 方法 | 信心 | 局限 |
|---|---|---|---|---|---|---|---|
| Intel Market Research LLM 中间件网关 | 2026-2034 | 全球 | 0.019 → 0.189 | 49.6% | 聚焦多 LLM 接入和管理的狭义中间件市场 | 低 | 可能是最窄的直接视角;方法不透明,并包含偏调查式的断言 |
| Intel Market Research AI API 网关 | 2026-2034 | 全球 | 0.85 → 2.12 | 12.0% | 更宽的 AI API 网关中间件,包含路由、转换、治理和分析 | 低 | 范围比 OpenRouter 更宽,因为它包含 API 中介和传统 API 网关功能 |
| The Business Research Company LLM 网关 | 2026-2030 | 全球 | 2.76 → 7.21 | 26.9% | 面向安全推理网关和 LLM 接入管理的软件、硬件和服务市场 | 中 | 包含硬件和服务,因此高估了仅托管路由器收入 |
| The Business Research Company LLM 网关平台 | 2026-2030 / 2035 | 全球 | 4.23 → 11.01 | 26.7%–27.0% | 横跨治理、流量与成本管理、监控的网关平台市场 | 中 | 同一页面在概览中把 11.01B 标成 2030 年结果,在报告属性框又标成 2035 年预测 |
| Research and Markets AI 网关 | 2025-2032 | 全球 | 3.66 → 9.61 | 14.7% | 覆盖智能网关解决方案及相关软件 / 服务的宽口径 AI 网关预测 | 低 | 范围看起来比 OpenRouter 的开发者路由器细分定位更宽,也更偏硬件 / IoT |
| 证据受限的 OpenRouter 相关 SAM(作者估算) | 2026 | 全球 | 0.20 → 1.00 | n/a | 托管式多模型路由 SaaS 切片,上界受 AI API 网关估算约束,下界受宽口径网关平台收入约束;排除自托管和硬件较重的细分 | 低 | 没有公开来源单独拆出 OpenRouter 的精确托管路由器细分 |
| 证据受限的独立路由器 SOM(作者估算) | 2026-2030 | 全球 | 0.03 → 0.15 | n/a | 上述 SAM 中,在云原生、API 管理和自托管替代品之后,独立托管路由器可能拿到的份额 | 低 | 对企业治理要求和云供应商捆绑强度高度敏感 |
该表刻意保留不兼容的公开视角,而不是取平均;最后两行是低信心作者估算,由已发布的窄口径和宽口径类别框定。
[CM007, CM008, CM009, CM010, CM011, CM012]OpenRouter 的现实市场从宽泛网关和控制平面收入,收窄到小得多的托管路由 SaaS 切片,再收窄到更小、可由独立厂商赢下的池子。
[CM012, CM015, CM017, CM048, CM049, CM050]已发布估算从窄口径 LLM 中间件延伸到数十亿美元的网关平台类别,因此应将品类表述为区间,而不是单一 TAM 数字。
最后一行保留 TBRC 页面内部不一致:概览文本称 2030 年达到 11.01B,而报告属性框将 11.01B 标为 2035 年预测。
[CM008, CM009, CM010, CM011, CM012, CM014]2.3 买方、用户、付费方与采用路径
OpenRouter 的第一批用户通常是开发者、产品工程师或 AI 构建者,他们想用最快路径试验模型,不想管理一堆 SDK 和供应商合同。用量扩大后,买方变宽。当公司需要配额、可观测性、故障转移、路由政策或跨团队标准化访问时,平台工程或 AI 基础设施负责人会变重要。在更大组织里,一旦 token 支出、部门级配额和 chargeback 变得重要,财务或 IT 治理就会成为共同付费方。这个买方地图也解释了分层。开发者主导的初创公司和 AI 产品团队适合托管路由器,因为它们看重速度、广度和轻运维。相反,受监管企业、集中式 API 平台团队,或必须混合内部与外部模型的组织,往往会评估完整 AI gateway 或自托管层。因此,采用路径从模型试验开始,走向团队级标准化;当 AI 用量从单一应用功能变成共享基础设施后,再走向集中治理。[CM026, CM032, CM033, CM034, CM036, CM037]
| 细分 | 主要买方 | 主要用户 | 付款方 / 预算负责人 | 工作流 | 采用触发点 | OpenRouter 为什么重要 |
|---|---|---|---|---|---|---|
| 初创公司产品团队 | CTO、工程负责人或产品负责人 | 应用开发者与 AI 工程师 | 产品工程预算或创始人控制支出 | 快速测试模型、迭代提示词、切换供应商 | 需要快速发布 AI 功能,又不想接入多个供应商 | 单一 API 和统一计费压缩首次生产实验时间 |
| 开发者平台 / API 团队 | 平台负责人或开发者基础设施负责人 | 平台工程师与内部工具开发者 | 共享平台预算 | 为多个内部应用标准化模型访问 | 供应商蔓延,或团队间反复做 API 接入 | 网关抽象减少重复接入,并形成一个内部访问层 |
| AI / ML 基础设施团队 | AI 平台 VP 或总监 | AI 工程师、基础设施工程师、模型运维人员 | 中央 AI 平台预算 | 按成本、延迟和可靠性在供应商之间分流工作负载 | Token 支出变得可观,模型基准测试也转为持续动作 | 路由、缓存和分析能把模型选择变成可衡量的控制闭环 |
| 合规敏感型企业 | CIO、安全或企业架构负责人 | 平台、安全和治理团队 | 中央 IT、安全和业务单元 AI 预算 | 在 AI 流量上统一执行政策、数据驻留和审计要求 | 受监管数据、数据区域边界或严格身份控制 | OpenRouter 能在实验层帮上忙,但在这一客群里往往输给自托管或企业级网关 |
| 代理机构 / 多租户 AI 构建商 | 技术创始人或交付负责人 | 解决方案工程师和客户侧构建人员 | 项目或客户交付预算 | 跨不同模型和用例承接大量客户提示词 | 需要成本归因、故障切换,并快速接入长尾模型 | 托管路由和统一账单减少多客户工作负载的运营开销 |
预算所有权通常先在工程团队手里,但 AI 用量变成共享基础设施、Token 支出需要配额或分摊后,会转向平台和治理团队。
[CM026, CM032, CM033, CM034, CM036, CM037]买方匹配度不只取决于模型广度,更取决于组织优先考虑速度、治理深度还是部署控制。
[CM032, CM033, CM034, CM036, CM037, CM039]2.4 增长驱动、ROI 逻辑与采用约束
路由层需求逻辑很直接。模型目录持续扩张,价格和质量按任务差异很大,供应商故障或速率限制 会打断面向客户的应用。路由、缓存和集中分析把这些问题变成明确的 ROI 杠杆。IDC 用性能、成本和信任来框定该类别;GetMaxim 认为混合工作负载路由可削减 40% 到 70% 的 token 支出;Google 自家模型阵容也说明,买方不想让每个 prompt 都走一个默认模型。但采用约束同样真实。托管 SaaS 路由器在 VPC 部署、内部密钥控制、审计轨迹和策略执行上弱于自托管或企业 AI gateway。原生云和现有 API gateway 已经为许多企业解决了足够多问题,这会限制独立路由器渗透。即便采用 gateway,团队也必须盯住代理延迟、精确匹配缓存限制,以及治理要求超过开发者优先托管服务能力的风险。[CM005, CM006, CM020, CM021, CM023, CM025]
| 驱动因素 / 约束 | 方向 | 时间 | 含义 | 尽调问题 |
|---|---|---|---|---|
| 模型增多与供应商更替 | 正向 | 当前 | 更多模型和供应商会推高对统一路由层的需求,企业不必再做一套套供应商专属集成 | OpenRouter 添加、下架或重排供应商的节奏,相对客户需求有多快? |
| 模型间价格分化 | 正向 | 当前 | 可接受模型之间成本差异很大,路由和缓存因此具备经济意义 | 客户流量中,自动路由相较固定模型调用占比多少? |
| 宕机、限流和故障切换需求 | 正向 | 当前 | 可靠性要求把团队推向后备链和多供应商韧性 | 企业买家在成交中提到故障切换或速率限制痛点的频率有多高? |
| 治理、信任和数据驻留要求 | 混合 | 当前至中期 | 这些要求会带来网关需求,但往往更偏向企业级或自托管方案,而不是开发者优先的 SaaS 路由器 | 除欧盟路由和供应商过滤外,OpenRouter 在企业合同里实际交付哪些控制? |
| 原生云和 API 网关替代品 | 负向 | 当前 | AWS、Microsoft、Cloudflare、Kong 以及内部平台团队可以吸收大部分同类支出 | 目标客户中,有多大比例已在现有技术栈里嵌入可信替代方案? |
| 仅 SaaS 架构和治理深度缺口 | 负向 | 当前至中期 | 受监管或集中化企业可能拒绝缺少 VPC 部署、完整 RBAC 或深度审计支持的托管路由器 | 多少后期交易因部署模式或合规原因流失? |
增长来自模型蔓延、价格分化和可靠性需求;采用受治理深度、替代品强度和部署模式要求约束。
[CM005, CM006, CM020, CM021, CM023, CM025]采用通常从实验和模型供应商蔓延进入路由,只有当 token 支出和可靠性成为共同问题后,才会走向集中治理。
[CM005, CM026, CM043, CM044, CM045, CM046]2.5 证据要点
03竞争格局
3.1 竞争版图与解决方案类型
OpenRouter 面对的不是一个干净的同业集合。市场至少分成五种替代路径来完成同一个买方任务。第一类是 LiteLLM 和 Martian 这样的中立路由层,承诺模型选择、故障转移或控制平面抽象,但不一定拥有底层算力。第二类是 Together AI、Replicate 和 Fireworks 这样的推理云,销售模型访问,也打包模型托管、训练或专用基础设施。第三类是 AWS Bedrock 和 Microsoft Foundry 内的云巨头路由,路由被打包进更宽的安全、采购和治理栈。第四类是 OpenAI 和 Anthropic 的直接供应商 API,定价透明,但把路由逻辑推回给客户。第五类是用 LangChain 和 LlamaIndex 等框架内部自建。这种结构很重要,因为 OpenRouter 通常是在和一种栈选择竞争,而不只是和另一个创业公司标识竞争。[CP008, CP013, CP017, CP019, CP023, CP026]
| 方案 / 类别 | 类别 | 规模 / 融资信号 | 目标客户 | 产品范围 | 定价信号 | 战略角度 / 局限 |
|---|---|---|---|---|---|---|
| OpenRouter | 中立路由层 | 官方文档强调供应商路由、后备、服务层级和专用路由器,而不是自有模型供给 | 想在上游供应商之间保留可携带性的应用构建者和平台团队 | 基于外部模型的多供应商路由、故障切换和专用路由逻辑 | 基于用量的路由,带 flex / priority 控制;实际企业折扣未公开 | 抽象层强;但不拥有模型供给,独占性较弱 |
| Together AI | 推理云 | 首页宣传 Series C 和 GPU 集群扩张 | 希望获得开放模型推理以及更深基础设施选项的开发者和初创公司 | Serverless、批处理、专用端点、专用容器和 GPU 集群 | Serverless 按 token 计费且无最低消费;专用端点面向更稳定需求;部分批处理折扣 | 基础设施范围广,但路由中立性不是叙事中心 |
| Replicate | 模型运行时 / 市场型替代 | 首页突出热门模型的数百万次运行记录 | 希望以最低运维快速发布模型驱动功能的开发者 | 通过一个 API 运行模型、微调并部署自定义模型 | 视模型而定,按硬件时间或输入 / 输出 token 计费 | 自助动作快;对跨供应商控制逻辑着墨较少 |
| Fireworks AI | 推理和训练平台 | 首页称每天处理 30T+ token | 希望在同一供应商获得开放模型推理、专用部署和训练的团队 | Serverless 推理、按需部署和训练流程 | Serverless 按 token 计费,按需部署按 GPU 秒计费 | 基础设施掌控深、性能定位强;中立性不如纯路由层 |
| Martian | 路由 + 合规套件 | 官方材料突出 Accenture 合作和 Airlock 合规发布 | 有模型治理顾虑的受监管或企业买家 | 模型路由,加上合规自动化和以可解释性驱动的优化 | 审阅来源未显示公开标价和客户数量披露 | 在受监管账户中可能有差异化,但公开商业透明度薄 |
| LiteLLM | 开源网关 | GitHub API 显示约 52k 星和约 9k fork;首页称 YC 支持 | 希望自托管或本地部署模型访问控制的平台团队 | 统一 OpenAI 格式访问、后备、预算、负载均衡和管理员功能 | 开源基础,叠加询价 / 本地部署销售动作 | 替代风险强,因为买方可以自托管并保留可携带性 |
| AWS Bedrock / Azure Model Router(云厂商路由) | 既有云平台路由器 | AWS 称有 100k+ 组织;Microsoft 将路由打包进更广泛的 AI 平台 | 已在 AWS 或 Azure 标准化的企业 | 更广泛云 AI 平台内的路由,叠加治理、安全和采购杠杆 | 云供应商用量定价和企业承诺,而不是简单的中立路由器标价 | 信任和渠道杠杆最强;路由范围比 OpenRouter 更窄,也更不中立 |
| 直连 API + 内部自建 | 现状替代方案 | OpenAI 和 Anthropic 披露直接定价,LangChain 和 LlamaIndex 提供路由器原语 | 有足够平台工程能力自主管理编排的团队 | 直接供应商集成,加上自定义路由、重试和中间件 | 透明的用量 API 定价,加上内部工程成本 | 控制力最高,对路由器供应商依赖最低,但实施负担最重 |
截至 2026 年运行日期,各行比较买方完成路由或模型访问任务的主要路径;公开融资或价格细节缺失时,表格采用最强可支撑的规模或分发信号,而不是编造数字。
[CP008, CP009, CP013, CP014, CP016, CP017]按路由中立性与分发或信任杠杆,对主要选项类别做序数地图。
坐标轴是基于所审阅来源集推导的序数综合评分,不是供应商报告的基准或市场份额数据。
[CP023, CP026, CP030, CP037, CP041, CP043]3.2 同业画像、能力与定价
在直接和相邻同业中,最重要差别是供应商卖的是路由中立性,还是更宽的基础设施所有权。买方明确想要多供应商可用性、故障转移和模型选择逻辑,又不想把工作负载迁到单一模型主机时,OpenRouter 最强。LiteLLM 和 Martian 从不同角度攻同一编排层:LiteLLM 靠开源平台工具,Martian 靠企业路由加合规自动化。Together 和 Fireworks 的竞争方式不同。两者都给开发者自助用量,但会把更大或更稳定的需求引向更深的基础设施足迹,包括专用 endpoints、按需部署或训练。Replicate 更偏运行时和市场替代,适合快速交付模型驱动功能,但与中立路由器的直接等同性更低。定价透明度不均:直接 API、Replicate、Fireworks 和 Together serverless 展示的用量经济性,比 Martian 或云企业承诺更可见。[CP004, CP005, CP008, CP009, CP010, CP011]
| 采购标准 | OpenRouter | Together / Fireworks | Replicate | Martian / LiteLLM | AWS / Azure 路由器 | 直连 API + 内部自建 |
|---|---|---|---|---|---|---|
| 跨供应商路由深度 | 高 | 中 | 低 | 高 | 中 | 买方自建 |
| 自动故障切换 | 高 | 部分 | 低 | 高 | 中 | 买方自建 |
| 自有基础设施或训练 | 低 | 高 | 中 | 低 | 高 | 低 |
| 公开自助定价清晰度 | 中 | 中 | 高 | 低至中 | 低至中 | 高 |
| 企业信任 / 合规边界 | 中 | 中 | 中 | Martian 高 / LiteLLM 中 | 高 | 取决于买方实现 |
| 自托管或本地部署路径 | 低 | 低至中 | 低 | 高 | 低 | 高 |
| 跨上游供应商中立性 | 高 | 中 | 低 | 高 | 低至中 | 高 |
单元格按采购标准总结,并非逐项功能完全等同。“买方自建”指能力可通过直连 API 和框架获得,但客户必须自行组装并运营。
[CP001, CP003, CP008, CP013, CP017, CP019]| 方案 / 类别 | 公开定价信号 | 合同模式 | 包含能力 | 未知项 / 限制 | 含义 |
|---|---|---|---|---|---|
| OpenRouter | 服务层级展示 flex / priority 控制,但广义企业经济性未公开 | 基于用量的 API,叠加企业条款 | 供应商路由、故障切换和专用路由器 | 实际企业折扣和赢单率数据未公开 | 技术试用容易,大规模商业基准更难做 |
| Together AI | Serverless 按 token 定价且无最低消费;批处理折扣和专用端点 | 可变用量走自助,较稳定或预留工作负载走销售辅助 | 开放模型访问,加上更深基础设施选项 | 专用定价和实际折扣未公开 | 对可能从路由升级到托管基础设施的团队有吸引力 |
| Replicate | 清晰的按需付费表述和模型特定价格估算 | 自助用量计费 | 模型执行、微调和自定义部署 | 未定位成中立路由合同 | 适合实验和发布功能,对可用性抽象的直接性较弱 |
| Fireworks AI | Serverless 按 token 计费,按需部署按 GPU 秒计费 | 自助加企业部署 | 推理、训练和专用部署 | 企业价目表和批量折扣未公开 | 当性能和基础设施深度比中立性更重要时形成竞争 |
| Martian / LiteLLM | LiteLLM 显示询价和本地部署动作;Martian 未显示公开定价 | LiteLLM 为开源加企业支持;Martian 为顾问式企业销售 | 网关控制、路由、合规或自托管杠杆 | 公开可比性有限,Martian 尤其如此 | 当买方想要控制权或受监管工作流定制时,这些方案更有吸引力 |
| AWS / Azure 路由器 | 定价嵌在云供应商模型定价、计算器和企业协议中 | 云消费加企业采购 | 路由叠加更广泛的平台治理和合规 | 仅路由经济性很难从更广泛云支出中干净拆出 | 当路由作为更大平台标准的一部分采购时,初创公司很难胜出 |
| OpenAI / Anthropic 直连 API + 内部自建 | 公开 token 表、服务层级和企业附加项可见 | 基于用量的 API,加上买方投入的内部工程 | 原始模型访问、供应商原生层级,以及自定义编排空间 | 买方必须在其上自建路由、故障切换和治理 | 给第三方路由器设定透明度底线,也让多归属在经济上仍可信 |
表格区分公开标价信号与不透明的企业承诺。公开来源未披露合同条款或实际折扣时,表格有意不暗示可以同口径比较的总拥有成本。
[CP004, CP005, CP010, CP011, CP012, CP015]按解决方案类别分组,展示谁掌握路由逻辑、基础设施、合规边界或自托管控制。
[CP023, CP024, CP030, CP037, CP038, CP040]3.3 巨头、开源与内部自建
最有力的替代品不一定是风投支持的同业。AWS 和 Microsoft 现在在更宽 AI 平台中营销路由,并把路由与治理、采购和信任姿态捆在一起,小型路由器很难匹配。Bedrock prompt routing 留在模型家族内,Azure model router 留在符合条件的已部署模型和 data-zone 规则内,所以二者都不如 OpenRouter 中立;但在既有云资产内,二者更容易被合理化。光谱另一端,直接 API 和开源框架让自建越来越可行。OpenAI 和 Anthropic 公开 token 定价和服务层级,LangChain 和 LlamaIndex 暴露路由或可移植性原语。LiteLLM 把这些世界接起来:它在 100+ 模型上标准化 OpenAI 格式,并加入 fallback、预算和负载均衡,实质性降低离开任何托管路由器的迁移摩擦。[CP023, CP024, CP026, CP027, CP028, CP029]
3.4 切换成本、分发力量与护城河耐久性
OpenRouter 的护城河真实但有条件。产品有可信的控制平面楔子——供应商负载均衡、故障转移、flex 或 priority 定价层,以及 Pareto 和 Fusion 等专用路由器——但护城河是软件逻辑,不是独家供给。因此,相邻基础设施厂商、hyperscaler 或开源 gateway 更容易复制足够多价值主张,从而压缩定价。最强分发优势属于 AWS 和 Microsoft,因为路由可以搭上更宽云合同和企业治理要求。最强切换压力来自 LiteLLM 加内部自建框架,因为这些工具保留模型可移植性,让多归属成本很低。Martian 的 RouterBench 逻辑也有双刃剑:它验证路由重要,也验证买方应在模型和供应商之间保持灵活。可能结果是一个结构性多归属类别,OpenRouter 可以靠便利性和速度赢单,但可能难以拥有整个栈或形成持久锁定。关键尽调测试是,溢价路由功能能否转化为可衡量的商业耐久性:多供应商客户留存更高,客户愿意为 Pareto 或 Fusion 额外付费,以及当买方可以默认选择捆绑云或自托管 gateway 时,赢率仍能守住。没有这些证明,路由便利性仍有价值,但经济上可被替代。[CP001, CP003, CP004, CP005, CP006, CP007]
| 护城河主张 | 主要威胁 | 严重度 | 当前证据 | 缓释措施 / 尽调问题 |
|---|---|---|---|---|
| 供应商抽象难以复制 | 开源网关和框架路由器已能复制足够多控制平面 | 高 | LiteLLM、LangChain 和 LlamaIndex 已提供后备、可携带性和路由器原语 | 要求按集成深度拆分同期群留存,并证明托管路由比自建更省工程时间 |
| 故障切换和可用性支撑溢价 | 云巨头和直连 API 不断加入自有路由和服务层级控制 | 高 | Bedrock 提示词路由、Azure 模型路由器和 OpenAI 式服务层级都会削弱托管路由的新鲜度 | 要求提供专门对云原生路由和直连 API 自建的赢单 / 输单数据 |
| 专用路由器创造差异化价值 | 竞争对手可以快速加入自己的判别、评分或策略层 | 中 | OpenRouter 推出 Pareto 和 Fusion,但护城河来自算法,而非独占供给 | 验证 Pareto / Fusion 功能的附加率和实际支付意愿 |
| 中立性胜过打包的既有巨头 | AWS 和 Microsoft 可以在大得多的合同里销售路由,并提供更强治理保证 | 高 | 云文档强调合规、治理和平台广度,初创公司很难照做 | 测试 OpenRouter 在受监管或采购繁重账户中是否仍能赢 |
| 品类受益于模型碎片化 | 多归属需求也会降低任一单一路由器的锁定 | 中 | RouterBench 认为,在成本、能力和延迟之间,没有单一模型始终最优 | 衡量客户是替换或补充 OpenRouter,还是排他性地标准化到 OpenRouter |
严重度反映对定价权和续约耐久性的压力,不代表路由这个品类是否会存在。风险登记表有意围绕可能压缩 OpenRouter 抽成率或渠道杠杆的证据来写。
[CP022, CP037, CP040, CP041, CP042, CP043]浓缩展示当前最强的几股力量:哪些在帮 OpenRouter 守住防御性,哪些在拖累。
取值是基于所引来源集合综合判断,而非外部评分卡。
[CP040, CP041, CP042, CP043, CP045]3.5 证据要点
04财务
4.1 收入模式、定价机制与确认口径注意事项
OpenRouter 的公开资料支持一种混合路由平台收入模式,但无法给出已实现收入的清晰拆分。官方来源显示,公司围绕用量变现,而不是披露年度合同:客户可以预先购买积分,通过 OpenRouter 的统一 API 路由推理,也可以选择把 BYOK 流量跑在同一控制平面上。最具体的官方费用不是按席位 SaaS 费,而是支付层经济性。OpenRouter 称,非加密货币充值收取 5.5% 费用,最低 $0.80;加密货币充值收取 5.0% 固定费;BYOK 只有在每月前 1 million 次请求之后,才成为 5% 的变现杠杆。Stripe Projects 显示免费默认计划和按量付费升级路径;Sacra 2025 年的研究估计,$100 million GMV 只对应 $5 million 年化收入,意味着大部分总支付量很可能流向上游供应商,而不是留在 OpenRouter 损益表上。反过来,这带来真实的收入确认问题:积分在消耗前购买,已用积分和已购积分分开跟踪,公开记录没有解释报道收入是总积分销售额、净抽成率,还是某种递延收入式余额。收入路径看得见,会计政策看不见。[CI004, CI008, CI009, CI016, CI020, CI025]
| 收入来源 | 机制 | 单位 | 当前价值 / 状态 | 质量 | 尽调问题 |
|---|---|---|---|---|---|
| 预付费路由推理 | 客户预购点数,用于 OpenRouter 跨供应商路由的非 BYOK 请求 | 用量驱动的点数 | 明显活跃且处于核心位置,但公开来源未说明收入是按总额确认,还是扣除供应商成本后按净额确认 | 中 | 提供按产品线拆分的收入确认政策和总额 / 净额列示 |
| 购买点数平台费 | 用户充值点数时,OpenRouter 留存手续费收入 | 充值金额的 % | 官方口径为非加密 5.5%,最低 $0.80;加密支付统一 5.0% | 高 | 展示手续费收入占已确认总收入的比例 |
| BYOK 超额费 | 客户自有供应商流量超过免费门槛后,OpenRouter 开始收费 | 等价请求金额的 % | 每月前 1M 次请求免费;超出后标准费率 5% | 高 | 拆出 BYOK 请求量、超额渗透率和实际抽成率 |
| 企业控制 / 支持 | 较大账户围绕路由层购买治理、工作区、预算、分析和协商支持 | 合同 / 订阅 / 服务 | 产品功能面可见,但定价和附加率未披露 | 低 | 提供企业包结构、ACV 区间和支持义务 |
| 免费层作为获客漏斗 | OpenRouter 补贴部分免费模型访问,在转化前吸引开发者 | 非收入获客支出 | 官方显示该模式活跃,且至少部分情况下由公司出资 | 中 | 量化免费到付费转化和每个激活账户的补贴成本 |
| 潜在合作伙伴 / 嵌入式渠道用量 | Stripe Projects 和合作伙伴分发可在无需单独直销的情况下带来用量 | 嵌入式用量 | 文档可见,但对签约额的贡献未知 | 低 | 拆出有多少用量来自嵌入式或合作伙伴主导流程 |
各行区分真正收入流与获客或渠道机制;公开证据能证明这些层存在,但看不到占比组合。
[CI004, CI008, CI009, CI020, CI021, CI025]| 来源 / 方案信号 | 价格 / 单位 / 合同 | 标价与实际 | 包含能力 | 折扣 / 未知项 | 含义 |
|---|---|---|---|---|---|
| 官方平台费公告 | 非加密充值收取 5.5%,最低 $0.80 | 标价规则 | 购买额度与钱包充值 | 未披露不同客户规模的实际组合 | 收入部分来自支付层经济性,不只是 token 转售 |
| 官方平台费公告 | 加密充值统一收取 5.0% | 标价规则 | 购买额度与钱包充值 | 加密充值占比未披露 | 对加密充值金额较大的账户,可能小幅改善贡献利润率 |
| 官方 BYOK 公告 | 每月前 1M 次 BYOK 请求免费,此后收取 5% 费用 | 标价规则 | 客户自有提供商密钥经由 OpenRouter 路由 | 实际请求结构和超额发生率未披露 | BYOK 能扩大采用,同时仍可从规模化用量变现 |
| 官方服务层级 | Flex 层级可比默认层级便宜 50%,代价是更高延迟 / 更低可用性 | 用量规则 | 路由层级选择与实际服务层级计费 | 实际层级组合未披露 | 单 token 平均收入取决于层级选择,不只取决于模型选择 |
| 独立方案梳理 | 免费层每分钟 20 次请求;充值 $10 后每日限额更高;按量付费可用 300+ 个模型 | 标价信号 | 试用与付费模型访问 | 来自第三方方案评测,不是结构化官方价目表 | 支撑低摩擦 PLG 获客,但看不到付费转化 |
| 企业版报价 | 面向支持、治理和 SLA 需求定制定价 | 协商价 | 工作区、访问控制和大客户支持 | 无公开价目表或最低承诺 | 企业变现可能存在,但无法用公开证据建模 |
公开定价主要是费率规则和定性方案描述;实际价格、折扣和企业合同结构未披露。
[CI008, CI009, CI011, CI012, CI027, CI028]OpenRouter 把开发者试用转化为由积分驱动的推理、费用抽取和企业扩张, BYOK 则是一条并行变现分支。
[CI008, CI009, CI016, CI020, CI033, CI038]4.2 GTM 路径、公开牵引力与销售效率代理指标
GTM 路径看起来获客阶段是产品驱动,扩张阶段逐步转向企业驱动。OpenRouter 官方和合作伙伴材料强调一个 API key、兼容 OpenAI 的 endpoint、支出可见性、路由权限、护栏、workspaces 和 analytics;这些都降低初始集成摩擦,让产品在采购前就容易测试。公开牵引力故事在广度上很强:OpenRouter 称其服务超过 8 million 用户或开发者,覆盖超过 400 个模型,并在六个月内把每周 token 量从 5 trillion 扩到 25 trillion。Stripe Projects 和免费层强化了漏斗从自助开始。Workspaces、analytics、用户活动 endpoints 和预算 API 显示产品正进入多团队或组织级使用,这是公开资料里最接近扩张收入的代理指标。缺失的是实际效率数学。公开资料没有 CAC、平均合同价值、赢率、净留存或回本期披露。用户数也没有拆成付费客户、企业组织或仅注册开发者。独立评论来源进一步表明,免费层 上限、支持响应和治理缺口,会在自主 agent 或生产工作负载扩张后变成转化摩擦。正确解读是,OpenRouter 的顶层漏斗异常强,但变现 cohort 的质量仍不透明。[CI002, CI003, CI017, CI018, CI019, CI020]
| 指标 | 数值 / 公开代理指标 | 置信度 | 重要性 | 尽调问题 |
|---|---|---|---|---|
| 年化收入 / GMV 代理指标 | 2025 年 5 月 $100M GMV 对应 $5M 年化收入(Sacra 估计) | 中 | 唯一类似收入的公开数据点,可衡量变现效率 | 提供当前 ARR、GAAP 收入以及提供商转付拆分 |
| 平台费变现 | 非加密充值费率 5.5%、最低 $0.80;加密充值 5.0% | 高 | 即使 token 价格只是转付,也能看到部分抽成率 | 拆分手续费收入与用量驱动收入 |
| BYOK 变现 | 每月 1M 次请求免费,此后收取 5% 费用 | 高 | 即便提供商对算力计费,OpenRouter 仍能从流量变现 | 提供 BYOK 请求量、超额发生率和毛利润 |
| 公开牵引力代理指标 | 8M+ 用户 / 开发者、400+ 个模型、每周 25T token | 高 | 指向漏斗顶部规模和使用密度 | 将用户拆分为付费账户、组织和企业 logo |
| 低余额运营阈值 | 建议余额 $10-20,以避免额外额度检查和延迟 | 中 | 暗示钱包经济性,也说明极小余额会抬高服务成本 | 展示低频充值或小额测试后流失的账户占比 |
| CAC / 回收期 / NRR / ACV | 低 | 公开材料缺少核心销售效率指标 | 提供 CAC 回收期、分客群 ACV、总留存和 NRR | |
| 毛利率 | 低 | 要判断 OpenRouter 更像 SaaS、支付,还是低毛利经纪模式,必须有这个指标 | 提供扣除提供商分成、支持成本和免费层补贴后的毛利率 |
空值表示确有披露缺口。可用的公开代理指标更多体现采用情况和费率规则,而不是实际软件经济性。
[CI002, CI003, CI008, CI009, CI013, CI014]公开的单位经济叙事从用量增长开始,经费用抽取和提供商分成, 最终通向一条尚未披露的毛利润线。
这座桥是概念性的,因为 OpenRouter 披露了费用规则和用量增长, 但没有披露实际收入组合、提供商分成比例或毛利率。
[CI009, CI011, CI012, CI016, CI021, CI025]4.3 成本结构、营运资本与资本充足性
OpenRouter 的成本结构在硬件资本开支上应明显轻于模型实验室,但这不等于一个简单的软件利润率故事。公司自己的文档把平台框定为 Cloudflare Workers、边缘缓存、供应商路由、回退逻辑、预算和 analytics,而不是自有 GPU 或基础模型训练。这指向低于推理供应商或实验室的资本强度。同时,非 BYOK 流量看起来仍会穿透上游模型成本,意味着毛利率取决于 OpenRouter 能通过平台费、BYOK 超额、企业支持和其他未披露合同经济性留下多少。可靠性也有真实交付成本:低余额会触发额外数据库检查,主路由失败会增加延迟,公司称其补贴部分免费模型 容量,2026 年 2 月故障还迫使团队做熔断器和回退缓存 工作。营运资本可能有利,因为客户先预付积分再消耗;但同一个钱包模式也带来递延服务和总额法 / 净额法判断。$113 million B 轮和约 $1.3 billion 估值信号之后,资本充足性方向上较强,但申报记录稀疏且滞后。可见 SEC 材料是一份 2025 年 Form D,只显示部分售出的豁免发行且没有收入披露,不是投资人完整承销需要的当前现金、现金消耗或资金续航。[CI001, CI005, CI006, CI007, CI012, CI013]
| 资本项目 | 公开数值 / 状态 | 置信度 | 重要性 | 尽调问题 |
|---|---|---|---|---|
| 2026 年 Series B | 2026 年 5 月宣布融资 $113M | 高 | 大额新股权融资支撑继续投入基础设施和企业能力 | 确认扣除费用后的现金到款和交割条件 |
| 投后估值信号 | TechCrunch 援引 NYT 报道称约 $1.3B | 中 | 给出投资人似乎愿意承销多少融资空间的参照 | 提供董事会批准的投后估值和完全稀释后股权 |
| 计划资金用途 | 基础设施扩容、企业能力、路由、治理和优化 | 高 | 说明可靠性和企业工具仍是支出重点 | 按产品、基础设施和 GTM 提供未来 12-24 个月支出计划 |
| 可见 SEC 豁免发行披露 | 2025 年 Form D 显示发行 $47.6M、售出 $29.6M,首次销售为 2025 年 5 月 27 日,30 名投资者 | 高 | 展示一次历史融资事件,但未呈现当前轮次的完整经济条款 | 将 Form D 线索与更完整的私募融资时间线对齐 |
| 手头现金 | 低 | 没有现金余额,就无法把资本充足性换算成续航月数 | 提供最近月末现金和受限现金余额 | |
| 净烧钱 | 低 | 烧钱速度把融资额转化为生存时间,也决定下一轮融资紧迫性 | 提供月度净烧钱,以及可靠性和企业团队招聘后的烧钱驱动因素 | |
| 续航月数 | 低 | 评估快速增长基础设施业务的融资依赖时,这一指标很关键 | 提供基准、下行情境和拉伸情境的续航月数 | |
| 债务或其他融资义务 | 低 | 风险债、云承诺或预留容量会显著改变风险 | 确认是否存在债务、最低云用量承诺或容量预购义务 |
公开融资支持很强,但现金、烧钱、债务和续航完全未披露,这是资本充足性无法从方向性判断走向可承销结论的主因。
[CI001, CI005, CI006, CI007, CI024, CI040]公开财务记录的跨度,从第三方估算的 $5M 收入,到 $1.3B 投后估值信号, 再到 2026 年 $113M 融资事件。
估值数字来自 TechCrunch 引用 NYT 报道;收入数字是 Sacra 的估算, 而非公司披露。
[CI001, CI007, CI024, CI025, CI040]OpenRouter 结合新股权融资和客户预付积分,为路由软件、企业控制项、 提供商分成和可靠性工作供血,但公开资金续航仍未知。
[CI005, CI016, CI018, CI019, CI021, CI022]4.4 财务结论与尽调阻断项
财务结论在需求和融资上正面,但底层单位经济性仍只有中等置信度。OpenRouter 显然有真实市场拉力:token 快速增长、多供应商相关性,以及足够的投资人需求来完成一笔大型战略 B 轮。产品也有几条变现杠杆——积分购买费、BYOK 超额、企业控制,以及潜在溢价支持——这些可能随用量扩大而叠加。问题是,关键承销输出没有一个公开。没有披露 ARR,没有自助式与企业之间的公开收入结构,没有供应商成本穿透拆分,没有毛利率、现金消耗、资金续航、客户集中度披露,也没有留存数据。负面证据同样重要。官方故障复盘和独立评论显示,一旦生产工作负载超过试验阶段,支持质量、可靠性工程和治理缺口会直接影响收入质量。这不意味着 OpenRouter 没有吸引力;它意味着下一步尽调不是再看一篇新闻稿,而是向管理层要数据。在把公司承销为持久基础设施、而不是快速增长的经纪层之前,投资人应要求收入确认备忘录、当前 P&L、现金消耗和现金计划、企业合同原型、cohort 留存、头部客户暴露,以及维持免费模型在线的经济性。在此之前,OpenRouter 财务上有希望,但还没有完全透明。[CI029, CI030, CI031, CI034, CI036, CI038]
| 缺失指标 | 对承销判断的影响 | 具体尽调路径 |
|---|---|---|
| ARR / 确认收入增长 | 阻断估值倍数、回收期和增长效率承销判断 | 索取月度经常性 / 非经常性收入桥,以及过去 12 个月确认收入 |
| 总额法 vs 净额法收入确认政策 | 不清楚额度销售、提供商转付和钱包余额是否高估经济收入 | 索取会计备忘录和审计师对主要责任人 vs 代理人处理及递延收入的意见 |
| 毛利率和提供商 COGS | 无法判断该模式更像高毛利 SaaS,还是更薄的经纪经济性 | 索取按收入流拆分的毛利率和提供商支付占比 |
| 现金、烧钱和续航 | 尽管有大额 Series B,资本充足性判断仍不完整 | 索取最新现金余额、烧钱趋势和 24 个月运营计划 |
| 企业合同组合和实际定价 | 没有合同原型,就无法把企业价值获取与自助服务费用拆开 | 审阅近期企业订单、最低承诺、折扣和支持义务 |
| NRR / 流失 / cohort 扩张 | 用来检验工作区和治理功能是否真正带来持久扩张 | 索取 cohort 留存表和加购 / 降级瀑布图 |
| 客户集中度 | 路由层看起来覆盖很广,但仍可能依赖少数高消费客户 | 索取前 10 大客户收入占比,以及按提供商或模型家族拆分的支出集中度 |
| 免费层补贴经济性 | 支持免费模型可能促进获客;若转化弱,也会吃掉毛利 | 索取补贴支出、免费到付费转化,以及按获客 cohort 拆分的回收期 |
这些是最低限度缺失字段;有了它们,OpenRouter 才能从强需求故事走向可融资的运营模型。
[CI034, CI035, CI039, CI040, CI041, CI042]4.5 证据要点
05产品与技术
5.1 产品定义与模块范围
OpenRouter 的产品最好理解为插在应用或 agent 与众多上游模型供应商之间的路由和控制平面。开发者不必分别集成 OpenAI、Anthropic、Google、xAI 或开放模型主机,可以指向一个兼容 OpenAI 的 endpoint,并在不重写应用结构的情况下切换模型选择。快速入门文档也让商业包装更清楚:团队可以留在原始 REST 以获得最大控制,也可以采用轻量客户端 SDK 获得类型安全和分页,或在需要工具循环和状态管理时上移到 Agent SDK。从客户工作流看,产品不只是更便宜的 token 访问。它把评估、供应商切换和多模型试验压缩进一个运营界面。模块地图现在已经从文本推理延伸到模型发现、排名、服务器工具、结构化输出、专用 图像 API、语音和转录,以及 workspace 治理。这种宽度让 OpenRouter 更接近 AI 连接基础设施,而不是狭义代理,尽管已审阅公开材料仍强调托管云访问,而不是私有部署。[CE001, CE002, CE003, CE004, CE031, CE034]
| 模块 / 资产 | 主要用户 | 状态 / 成熟度 | 核心功能 | 差异化 | 尽调缺口 |
|---|---|---|---|---|---|
| 统一推理 API | 应用和 agent 开发者 | 成熟核心 | 一个兼容 OpenAI 的端点,接入数百个模型和提供商 | 省掉逐个提供商集成的工作,保持低切换成本 | 未公开按层级拆分的吞吐量或 SLO 基线 |
| 模型目录和排名 | 评估人员和有采购意识的工程师 | 成熟 | 按模态、价格、上下文和 benchmark 维度浏览、筛选、比较模型 | 把路由标准变成机器可读,而不是口口相传的经验 | 排名方法和反作弊控制未完全公开 |
| 路由引擎(Auto Router / Auto Exacto) | 平台工程师 | 成熟且在扩展 | 根据 prompt、健康状态和工具调用表现选择或重排模型与提供商 | 在商品化推理访问之上沉淀运营 know-how | 未公开每种路由模式按工作负载拆分的 benchmark 差异 |
| 可靠性和成本控制 | 生产环境运维方 | 成熟核心 | 响应缓存、自动故障转移、模型 fallback 和零 completion 保险 | 把 uptime 和重试经济性产品化 | 公开文档仍承认存在一些边缘计费路径 |
| 服务器端工具和结构化输出 | Agent 构建者 | Beta 与成熟能力并存 | 增加服务器端 web 工具、图像生成和 schema 安全输出 | 把编排逻辑移入控制平面 | 服务器端工具安全有效性指标未公开 |
| SDK 和包接口 | JavaScript 和 Python 开发者 | 活跃 | 官方 TS、Python、agent 和 AI SDK 提供商包 | 在既有生态和运行时模式里触达开发者 | 包接口可见,但企业级保障因框架而异 |
| 工作区治理、BYOK 和 ZDR | 企业管理员 | 成长中 | 支出控制、白名单、PII 脱敏、按提供商划分的 ZDR 和 BYOK 管理 | 在共享模型市场之上增加治理层 | 公开审计材料和正式 SLA 条款仍是待尽调问题 |
| 媒体和私有模型扩展 | 多模态产品团队 | 2026 年新增 | 专用图像 API、语音和转写、私有模型路由 | 将平台从纯文本撮合扩展出去 | 新模块采用情况、支持负载和收入组合未披露 |
各行总结官方文档、发布说明、包接口和独立集成参考中明确记录的不同产品界面。
[CE001, CE003, CE007, CE009, CE012, CE017]| 用户任务 | 当前工作流痛点 | OpenRouter 方案 | 公开收益信号 | 限制 |
|---|---|---|---|---|
| 快速比较前沿模型 | 每个提供商都有单独密钥、端点和 schema | 统一端点,加模型目录和排名 | 无需重构核心应用逻辑,就能试验和切换模型 | 底层模型质量和提供商特性仍有差异 |
| 构建使用工具的 agent | 提供商专属函数 schema 和脆弱编排循环 | 标准化工具调用,加 Agent SDK 循环和状态 | 共享工具和消息模式可跨模型复用 | 工具可靠性仍取决于 schema 质量和提供商行为 |
| 提供商出问题时保持应用在线 | 单一提供商宕机和 429 上限会变成用户可见错误 | 自动提供商故障转移、可选模型 fallback 和零 completion 保险 | 重试设计比直接集成更便宜,也更有韧性 | 过窄白名单会减少可用恢复路径 |
| 降低重复推理成本 | 回归测试和重复 prompt 会让同一工作重复计费 | OpenRouter 层响应缓存 | 相同且成功的请求可即时返回,且计费使用量为零 | 工具调用和错误响应不可缓存 |
| 为 agent 增加实时 web 上下文 | 团队必须自建或托管搜索和抓取工具 | 服务器端 web 搜索和 web fetch 工具 | 模型可在单次请求流程中决定何时调用实时信息工具 | 服务器端工具仍处 beta,需要治理控制 |
| 集中隐私和支出控制 | 每个上游提供商的日志和政策界面都不同 | 工作区护栏、ZDR 视图、提供商日志和 BYOK 管理 | 治理成为提供商集合之上的一层 | 公开保障证据比功能界面薄 |
工作流行聚焦产品在实践中解决的客户任务,而不只列营销功能名。
[CE001, CE004, CE012, CE013, CE015, CE019]用户流程从兼容应用请求进入路由策略,再到可选工具执行, 最后落到有日志记录的响应界面。
[CE001, CE002, CE015, CE017, CE019, CE020]5.2 路由架构与运营模式
架构上,OpenRouter 把标准化请求接口与政策和路由逻辑结合起来,运行在上游供应商之上。models API 暴露定价、模态、上下文、供应商、作者、benchmark 和 ZDR 过滤器,把模型选择标准变成机器可读输入,而不是临时的单供应商代码。Auto Router 随后加入 prompt 感知的模型选择、会话粘性,以及 allowed_models 和成本质量偏好等路由控制。Auto Exacto 是专门面向工具使用的第二层:它按吞吐量、benchmark 信号和实测 tool-call 成功率重排供应商,并明确跟踪 invalid JSON、unknown tool names 和 invalid arguments 等失败桶。可靠性功能位于模型执行前后两侧。请求到达任何供应商之前,响应缓存在 OpenRouter 层发生;供应商故障转移是自动的,模型回退则通过 models 数组显式给出。Tool calling、结构化输出和服务器工具把控制平面延伸到运行时编排,让团队能标准化 schema,并加入搜索或 fetch 能力,而不用自己构建每一个工具 wrapper。[CE005, CE006, CE007, CE008, CE009, CE010]
| 层 / 组件 | 角色 | 公开证据 | 关键依赖 | 主要风险 |
|---|---|---|---|---|
| 客户端应用和 agent 框架 | 产生 prompt、工具 schema 和会话上下文 | Quickstart 和外部框架文档 | 开发者 SDK 和 OpenAI 兼容抽象 | 兼容性破裂会很快在下游工具暴露 |
| OpenAI 兼容 API 和 SDK 层 | 跨提供商统一请求和响应格式 | Quickstart、TypeScript SDK、Python SDK 和 AI SDK 提供商文档 | Base URL 兼容性和包维护 | 浅层兼容性主张比深层控制平面行为更容易复制 |
| 模型目录和元数据服务 | 为选择逻辑暴露模型、提供商、上下文、价格和 benchmark 信息 | Models 文档和 get-models API 参考 | 准确的上游元数据和 benchmark 刷新 | 元数据陈旧会拉低路由决策质量 |
| 路由和优化引擎 | 执行 Auto Router、Auto Exacto、故障转移、fallback 和健康感知选择 | Auto Router、Auto Exacto、uptime 和故障转移文档 | 提供商健康遥测和策略配置 | 对提供商约束过多会削弱韧性 |
| 可靠性、缓存和计费控制 | 缓存重复请求,免除零 token 失败计费,并暴露使用信号 | 响应缓存和零 completion 保险文档 | 正确的缓存键和收费对账 | 边缘计费路径仍是已披露的保留项 |
| 服务器端工具执行层 | 代表模型运行搜索、抓取、图像和 panel 工具 | 服务器端工具概览及特定工具文档 | 工具执行安全、外部引擎和速率控制 | Beta 状态和安全透明度仍不完整 |
| 工作区治理和隐私控制 | 应用护栏、日志、ZDR、EU 路由和 BYOK 管理 | 提供商日志、输入 / 输出日志、ZDR preview 和 5 月发布说明 | 管理员配置质量和企业方案功能 | 公开保障材料落后于治理功能集 |
架构行反映有文档记录的控制平面层,而不是任何未披露的内部基础设施拓扑。
[CE005, CE007, CE008, CE010, CE012, CE015]这套架构把兼容层、路由逻辑、控制面元数据、治理和提供商执行放在一个 API 之下。
层次划分综合自官方文档和生态包接口,而不是内部架构图。
[CE001, CE003, CE005, CE007, CE012, CE017]OpenRouter 依赖提供商健康度、准确元数据、外部工具引擎和谨慎的工作区策略配置。
[CE005, CE017, CE024, CE026, CE027, CE028]5.3 差异化、生态与 2026 年路线图
OpenRouter 的差异化一部分来自技术,一部分来自生态。技术上,公司不断把原始推理之上的逻辑加厚:类型化模型与 endpoint 元数据、路由旋钮、tool-call 优化、成本保护、多模态能力发现,以及 workspace 级治理。2026 年产品信号是同一控制平面论点的扩展,而不是品类转向。专用图像 API、语音和转录、Model Fusion、私有模型路由、BYOK 管理和按供应商 ZDR 控制,都让平台更适合作为 agent 团队共享运营层。外部生态同样重要,因为它降低采用摩擦,也充当开发者信号。官方代码仓库和软件包 页面显示,TypeScript、Python、agent 和 AI SDK provider 包持续维护;LangChain、LiteLLM、OpenHands、Cline、AI SDK 和 Pydantic 的外部技术文档则显示,开发者可以通过已经用于 agents、流式处理、结构化输出和自定义 OpenAI 兼容客户端的工具消费 OpenRouter。这种广度是真实分发资产,但也意味着护城河很多来自路由经验 和运营表面,而不是独家模型 IP。[CE031, CE032, CE033, CE034, CE035, CE036]
| 日期 / 时段 | 功能 / 里程碑 | 状态 | 影响 | 来源视角 |
|---|---|---|---|---|
| 2026 | 专用 Image API | 已发布 | OpenRouter 进入多模态生成,并提供能力发现和标准化定价元数据 | Unified Image API 公告 |
| May 2026 | 语音与转写 API | 已发布 | 将同一套密钥和路由界面延伸到语音工作流 | 5 月发布亮点 |
| May 2026 | Model Fusion | 已发布 | 新增多模型综合,覆盖插件、服务器工具和聊天室功能 | 5 月发布亮点 |
| May 2026 | Private Models(Enterprise) | 已发布 | 客户可在同一控制平面路由专用或微调端点 | 5 月发布亮点 |
| May 2026 | BYOK 管理 API | 已发布 | 把密钥管理升级为核心企业管理界面 | 5 月发布亮点 |
| May 2026 | 按提供商配置的 ZDR 控制 | 已发布 | 隐私约束可按提供商区分,不必限制整个模型目录 | 5 月发布亮点 |
| May 2026 | session-id 粘性和 rankings-daily 数据集 | 已发布 | 改善 agent 工作流连续性,并让排名分析可编程 | 5 月发布亮点 |
本路线图表使用已发布的公开发布说明,而非未公开的内部计划。
[CE031, CE032, CE033, CE034, CE035, CE036]核心路由已经成熟,生态触达很广;主要成熟度缺口不在功能数量, 而在企业级保障。
[CE018, CE031, CE037, CE046, CE047, CE048]5.4 信任、隐私、支持与尽调缺口
信任和治理在产品里很可见,但风险没有完全解除。正面看,OpenRouter 发布了供应商级训练和留存披露、EU routing、ZDR 感知 endpoint 预览、支出和白名单控制、提示词注入阻断、PII 脱敏,以及仅管理员可见的输入输出日志,且使用加密存储。Zero completion insurance 和自动供应商故障转移也显示,成本保护和可用性 被当作产品功能,而不只是支持承诺。开放尽调项同样重要。日志留存和 EU routing 限制意味着,隐私行为仍取决于功能开关和流量路径。供应商留存偏好有披露,但路由政策不会自动强制执行。独立评论也提示生产阶段围绕支持速度和治理深度的担忧。最重要的是,已审阅公开材料没有验证公开 SLA、SOC 2 或 ISO 27001 等第三方审计包,也没有公开 VPC 或自托管部署架构。现有证据支持一个面向成熟开发者的可信云路由层,但还不能完全关闭企业保证和支持义务问题。[CE021, CE022, CE023, CE024, CE025, CE026]
| 控制项 / 指标 | 当前公开状态 | 范围 | 作用 | 剩余缺口 |
|---|---|---|---|---|
| 提供商训练退出设置 | 文档已说明,可配置 | 付费模型和免费模型有独立控制 | 避免把请求路由给可能用提示词训练的提供商 | 不会改变 OpenRouter 自身独立的数据使用设置 |
| 提供商留存可见性 | 文档已说明,但仅作提示 | 按提供商展示留存和训练政策 | 买方可在使用前筛选提供商 | 路由策略不会自动执行留存偏好 |
| 欧盟区内路由 | 文档已说明的企业功能 | 流量发送至 eu.openrouter.ai | 支持区域数据处理要求 | 公开文档未说明该模式的定价、SLA 或架构细节 |
| ZDR 端点预览及按提供商的 ZDR 控制 | 文档已说明 | 端点级元数据,加上按提供商配置的开关 | 路由前把隐私取舍摆明 | 公开文档未提供对提供商侧行为的独立验证 |
| 输入和输出日志 | 文档已说明的 beta 功能 | 仅管理员可见、加密存储、至少保留三个月 | 调试、评估和用量复查 | 欧盟路由流量会跳过该功能,删除仍需联系支持 |
| 工作区护栏 | 2026 年 5 月发布项已有文档说明 | 支出上限、允许名单、提示注入拦截和 PII 脱敏 | 在密钥和用户层集中安全与治理策略 | 未找到公开的误报率、召回率或效果指标 |
| 正式保障材料包 | 审阅来源中未见公开验证 | SLA、SOC 2 / ISO、渗透测试和私有部署证明 | 可显著降低企业尽调摩擦 | 仍需向管理层索取,并非公开材料 |
本表将可见的产品控制,与审阅公开材料中仍缺失的第三方保障证据分开。
[CE021, CE022, CE023, CE024, CE025, CE026]5.5 证据要点
06客户
6.1 客户分层、购买中心与核心工作负载
OpenRouter 的客户故事从开发者开始,但公开资料显示,这个标题下面至少有四种不同经济动线。官方材料反复描述一个面向开发者和企业的 API;workspace 和 Stripe Projects 文档则补足了随着用量成熟,买方、用户和付费方如何分离。个人构建者和初创团队看起来是最容易进入的顶层漏斗:他们可以快速开通访问、比较多个模型,并避免管理多个供应商 SDK 或密钥。下一段是嵌入式工具——Roo Code、OpenHands、Aider 和 LiteLLM 等编程 agents 与中间件供应商,它们明确记录支持 OpenRouter。这些工具很重要,因为它们既是客户也是渠道,不经过传统采购周期就把 OpenRouter 拉进编程工作流。再往上是组织买方。Workspaces、用户活动分析、隐私设置、服务层级和 workspace 预算都说明,OpenRouter 正试图把最初的自助动线转成多团队治理和支出管理。公开证据在地区和垂直结构上远比产品形态薄:最清楚的地理信号只是“全球用户”,最清楚的垂直信号则是公开证据集中在软件、agent 工具和其他 AI 原生 工作负载。因此,本章可以有信心地绘制买方、用户和付费方分层,但还无法量化按地域或行业划分的变现需求。[CU001, CU002, CU004, CU005, CU006, CU007]
| 客群 | 买方 / 用户 / 付款方 | 地域 / 垂直行业 | 典型用例 | 证据和规模信号 | 缺口 / 含义 |
|---|---|---|---|---|---|
| 个人开发者和独立构建者 | 自助模式下,买方 = 用户 = 付款方 | 全球;公开证据集中在软件场景 | 快速比较模型、试提示词、使用免费模型、一把密钥接入 | Series B 轮公告提到 8M+ 开发者;Aider 和 Stripe Projects 强调快速设置与免费计划试用 | 漏斗顶部很强,但公开证据未说明有多少转化为长期付费账户 |
| 初创公司产品和 AI 团队 | 工程负责人或创始人购买,开发者使用,公司卡或额度付款 | 明确提到 AI-native 初创公司;公开地域细节很少 | 无需接入多个提供商,就把多模型推理嵌入产品 | Business Wire 称 AI-native 初创公司在使用 OpenRouter;TrueFoundry 称这一阶段看重快速切换模型和降低密钥管理负担 | 初创公司需求真实存在,但收入分层不可见 |
| 嵌入式工具供应商和框架 | 平台维护者或产品团队购买;终端用户通过工具消费;终端客户或供应商可能付款 | 开发者工具垂直;具名证据主要来自编程和 agent 生态 | 为编程 agent、中间件、IDE 和开源框架路由 LLM 流量 | Roo Code、OpenHands、Aider 和 LiteLLM 都记录了对 OpenRouter 的一等支持 | 这是高信号采用证据,但也可能造成渠道依赖 |
| 多团队组织 | 工程、平台或 AI 运营负责人购买;多个团队使用;统一账单付款 | 跨团队内部 AI 部署;具体垂直行业组合未披露 | 在一个账户下分离团队、项目和部署阶段,并套用共同治理 | Workspaces 文档增加了组织角色、共享账单、用户分析和工作区控制 | 展示了扩张路径,但未说明实际有多少组织在使用 |
| 企业或受监管买方 | 采购、平台或安全买方付款;运维和应用团队使用 | 大型企业被笼统提及;按行业拆分的变现组合不公开 | 受治理的生产推理,带支出可见性、隐私控制和协商支持 | Business Wire 和 CapitalG 强调大型企业使用、路由权限和便于审计的报表 | 公开的 SLA、续约和集中度细节仍太少,无法支撑完整投资判断 |
| BYOK 高阶用户和受采购约束的团队 | 开发者或团队负责人购买提供商直连额度;组织支付提供商账单加 OpenRouter 费用 | 编程工具、图像应用和多提供商试验 | 保留与提供商的直接关系,同时把 OpenRouter 用作路由层 | Sacra 描述了 Cline、Aider 和 Fal.ai 中的 BYOK 使用;Roo Code 记录了 BYOK 定价处理 | 有助采用,但如果终端客户保留上游议价力,OpenRouter 的抽成会受限 |
各行拆分公开来源能支撑的自助、嵌入渠道和组织型获客路径;地域和垂直行业组合仍停留在宽泛描述,没有数字披露。
[CU001, CU002, CU004, CU005, CU006, CU008]展示 OpenRouter 的公开获客路径:从自助试验,到嵌入式工具, 再到多团队治理,而不是经典的自上而下企业销售。
[CU004, CU006, CU010, CU020, CU031, CU035]比较 OpenRouter 公开界面与不同客户细分市场的匹配度,突出便利性最强的位置, 以及治理或支持缺口开始变得更实质的位置。
细分市场匹配度评分是基于保留的公开文件做出的定性综合, 尤其参考工作区文档、定价分析、评论证据和故障披露。
[CU006, CU009, CU010, CU032, CU034, CU035]6.2 采用轨迹与具名客户证明
OpenRouter 客户最强的公开证据是用量增长的广度,而不是一长串企业案例。官方和第三方 B 轮材料在核心轨迹上高度一致:每周流量在六个月内从 5 trillion tokens 升到 25 trillion tokens,月度量达到约 100 trillion tokens,平台称覆盖超过 400 个模型、拥有超过 8 million 用户或开发者。Sacra 2025 年研究给了一个有用的早期锚点,估算公司从 2024 年底约 $19 million 年化支出和 $1 million 年化收入,跃升到 2025 年 5 月约 $100 million 年化 GMV 和 $5 million 年化收入。这是真实采用动能。资料中更薄的是具名证明。公开、直接客户证据主要由开发者工具文档主导,而不是经典企业 标杆客户案例。Roo Code、OpenHands、Aider 和 LiteLLM 都维护明确的 OpenRouter 集成页面;Product Hunt 归档评论还加入来自 involve.me、Clado 和 Agents Base 的具名创作者 证言。这些都是可信证据,说明 OpenRouter 确实嵌入活跃工作流,尤其是编程和自主 agent 工作流,而且至少有部分客户超过玩具实验阶段。但这仍不同于披露的 Fortune 500 生产部署,后者还应带有续约历史和合同经济性。正确解读是,OpenRouter 有丰富用量信号和不错的具名开发者工具证明,但直接企业客户证明比融资叙事更薄。[CU003, CU013, CU014, CU015, CU016, CU017]
| 指标 | 数值 / 证据 | 日期 | 来源质量 | 含义 | 缺失分母 |
|---|---|---|---|---|---|
| 年化支出代理指标 | 年化支出约 $19M | 2024-12 | Sacra 估算 | 说明在 2025–2026 融资周期前,平台已有可观的变现使用 | 不是公司披露的 GMV 或收入项 |
| 年化收入代理指标 | 年化收入约 $1M | 2024-12 | Sacra 估算 | 表明到 2024 年底,OpenRouter 已找到按抽成变现的商业模式 | 未披露经审计收入或毛利率 |
| 年化 GMV 代理指标 | 年化 GMV 约 $100M | 2025-05 | Sacra 估算 | 表明 Series B 轮前,开发者和嵌入式工具采用很快 | GMV 不说明客户数量或集中度 |
| 年化收入代理指标 | 年化收入约 $5M | 2025-05 | Sacra 估算 | 隐含变现随使用量快速上升 | 仍没有公司确认的 ARR 或 GAAP 收入 |
| 周 token 量 | 六个月内,周 token 量从 5T 增至 25T | 2025-11 至 2026-05 | 官方披露 + 独立媒体 | 明确信号:采用在 2026 年加速,工作负载更接近生产 | token 量不等于独立付费组织数 |
| 用户和模型广度 | 8M+ 用户或开发者、400+ 模型;月 token 量约 100T | 2026-05 | 官方披露 + 独立媒体 | 说明平台对多家模型提供商都有广泛价值 | 用户总数未拆分为免费、付费、企业或活跃组织 |
| 运营扩张界面 | Workspaces、按工作区预算、30 天活动分析、服务等级和 Stripe 配置 | 2026-07-01 | 官方文档 | 说明产品已为多团队扩张埋好工具,不只是自助试用 | 功能可用不等于已披露的客户渗透率 |
| 生态发现入口 | 公开的应用 / agent 和 LLM 排名 | 2026-07-01 | 官方站点入口 | 表明外部活动足够持续,才值得推出公开排名和发现功能 | 排名不披露收入或续约质量 |
该轨迹把已披露的运营指标、明确标注的 Sacra 估算和平台功能代理信号放在一起,因为 OpenRouter 不公布客户数、ARR 或付费组织队列。
[CU003, CU013, CU014, CU017, CU018, CU019]| 具名客户 / 工具 | 客群 | 部署 / 用例 | 生产 / 试点 | 结果或证据 | 局限 |
|---|---|---|---|---|---|
| Roo Code | 编程 agent / IDE 扩展 | 官方提供商集成,包含 API 密钥设置、100+ 模型、BYOK 和提示缓存指引 | 生产级支持集成 | Roo Code 维护专门的 OpenRouter 指南,并自动从 OpenRouter 拉取模型选项 | 公开文档能证明支持使用,但不能说明合同规模,也不能说明 Roo Code 是否为付费企业账户 |
| OpenHands | 云端编程 agent / 开发者平台 | OpenHands 内的 LLM 提供商选项,明确给出 OpenRouter 设置和推荐模型字符串 | 生产级支持集成 | OpenHands 将 OpenRouter 作为 Agent SDK 和界面的权威提供商参考写入文档 | 文档证明活跃支持,但不能证明支出、合同条款或客户留存 |
| Aider | CLI 编程 agent | 命令行用 `openrouter/<provider>/<model>` 字符串路由模型 | 生产级支持集成 | Aider 称许多用户通过 OpenRouter 使用 Sonnet,并强调可经由 OpenRouter 使用免费模型 | 证据来自社区和文档,而不是公开企业案例 |
| LiteLLM | 网关 / 中间件供应商 | LiteLLM 将 OpenRouter 列为生产环境变量配置和多模态调用的一等提供商 | 生产级支持集成 | LiteLLM 在文本、视觉和 embedding 用例中都把 OpenRouter 当作真实提供商选项 | 这更能证明生态嵌入,而不是终端客户经济性 |
| involve.me 创作者社区 | 构建者 / 产品团队 | Product Hunt 创作者评价摘要称,OpenRouter 有助于多模型测试和成本 / 质量优化 | 公开证言 | Product Hunt 评价摘要中出现了具名创作者的好评 | 评价摘要属于二手归档材料,未披露付费深度 |
| Clado 和 Agents Base 的创作者 | 构建者 / agent 产品团队 | Product Hunt 评价摘要称 OpenRouter 能可靠处理 LLM 流量,并快速接入新选项 | 公开证言 | 同一评价语料中出现多位具名创作者,说明公司之外也有人背书 | 创作者评价不等于已披露的生产合同 |
这是本章留用的具名公开客户证据的部分清单。它捕捉截至 2026-07-01 可见的最强具名集成和证言,不代表完整已部署客户基础。
[CU021, CU022, CU023, CU024, CU026, CU027]追踪一条有证据支撑的路径:公开用量增长如何进入有文档记录的部署和生态渠道, 并标出证据哪里强、哪里开始变薄。
[CU013, CU014, CU017, CU020, CU021, CU023]按部署成熟度、结果具体性、新鲜度,以及证据是直接合同证明还是受支持集成 / 证言, 评估公开具名客户证据的质量。
矩阵单元格定性评估证据质量。“当前”表示 2026-07-01 可直接访问; “历史”表示存档快照或较早的独立研究,对客户证据判断仍有参考价值。
[CU021, CU023, CU024, CU026, CU028, CU029]6.3 耐久性、扩张与采购摩擦
公开记录支持一个可信的先落地再扩张故事,但还不能给出耐久留存结论。正面看,OpenRouter 已经搭出真实组织在首次采用后会需要的界面:workspaces、按 workspace 的密钥、组织角色、统一账单、按用户活动分析、企业预算、服务层级控制,以及 opt-in 数据处理设置。这些是扩张功能,不是业余项目的小便利。它们说明 OpenRouter 想让客户从自助试验进入多团队部署。与此同时,缺失指标正是投资人承销客户质量前最想看的。公开资料没有 NRR、GRR、churn 披露,没有合同期限披露,也没有集中度披露。负面证据在这里同样重要。TrueFoundry 的评论和定价分析描述了早期构建者与生产买方之间的清晰分裂:OpenRouter 因统一模型访问和快速切换受到称赞,但也因支持响应、账户安全处理、工作流级支出控制、缺少公开 SLA 标准,以及组织进入受监管或高用量场景后出现的治理限制受到批评。OpenRouter 自己的 2026 年 2 月故障复盘强化了同一点。公司透明并修复了 401-versus-503 行为,但事故仍削弱大买方需要的可靠性承诺。扩张可信,但没有续约数据和集中度细节,耐久性仍未证明。[CU020, CU030, CU031, CU032, CU033, CU034]
| 指标 | 数值 / 公开代理指标 | 客群 | 置信度 | 含义 | 尽调问题 |
|---|---|---|---|---|---|
| 净收入留存 | null | 所有付费队列 | 低 | 没有公开 NRR,扩张质量无法验证 | 索取按直销企业、初创公司和嵌入渠道队列拆分的 NRR |
| 毛留存 / logo 流失 | null | 所有付费队列 | 低 | 未披露流失率,业务可能仍是高量但浅层 | 索取按季度拆分的 GRR、logo 留存和流失原因 |
| 平均合同期限 / 续约节奏 | null | 企业账户 | 低 | 仅凭融资或 token 量推不出合同耐久性 | 索取标准 MSA、订单、合同期限和续约日历 |
| 重复使用埋点 | 按端点分组的 30 天用户活动,加统一工作区账单 | 组织账户 | 中 | OpenRouter 已跟踪可支撑客户成功和支出复盘的行为 | 要求管理层公布或共享这些内部系统的队列汇总 |
| 公开满意度分化 | Product Hunt 归档评价显示 86 条评价得分 5.0/5;TrueFoundry 报告的 2026 年 5 月 Trustpilot 快照为 41 条评价得分 1.7/5 | 构建者和生产评价渠道 | 中 | 产品在创作者圈层口碑较好,但在支持压力更重的生产场景明显偏弱 | 索取按客群拆分的 NPS / CSAT,以及工单响应时间数据 |
| 免费层重复使用摩擦 | 免费模型使用限制为 20 req/min 和每日上限;智能体式或批处理工作负载没有工作流止损时,仍可能撞上 429 或超支 | 试验用户和低支出用户 | 中 | 免费转付费可能受便利性推动,也可能被配额摩擦或意外支出拖累 | 索取免费和低额度队列的转化、滥用和限流发生率 |
null 表示公开记录未披露该指标。表格有意将硬留存指标与可观察的使用或满意度代理指标分开。
[CU011, CU025, CU030, CU031, CU032, CU033]| 扩张驱动因素 | 集中度或摩擦风险 | 影响 | 尽调路径 |
|---|---|---|---|
| 自助开发者可通过工作区、共享账单和预算升级为团队账户 | 没有公开数据说明免费或初创账户能转化为耐久企业合同 | 增长面可能很宽,但质量低于 token 量标题暗示的水平 | 索取从首次购买额度到多工作区组织采用的队列转化 |
| Roo Code、OpenHands、Aider 和 LiteLLM 等嵌入式工具扩大分发 | 渠道集中度或对编程 agent 生态的依赖未披露 | 少数主要工具或框架一旦调整,可能拖慢新增需求和用量质量 | 索取按头部嵌入合作伙伴和产品渠道拆分的 GMV 与收入组合 |
| 预算、隐私设置、服务等级和分析等企业控制创造增购入口 | 预算仅限 Enterprise 计划,SLA 条款靠协商而非广泛公开 | 大买方在扩大支出前,可能要求控制能力或支持承诺 | 索取企业包附加率、SLA 模板和采购赢单 / 输单原因 |
| 即使客户直接持有提供商关系,BYOK 也把客户留在 OpenRouter 控制平面上 | BYOK 可能限制 OpenRouter 抽成,也让单看表层用量更难判断收入集中度 | 若更多流量走客户自有密钥,GMV 或 token 增长可能高估净收入质量 | 索取 BYOK 流量占比、费用抽成率,以及 BYOK 相对非 BYOK 账户的留存 |
| 公开排名和应用入口增强发现与可信度 | 发现入口不说明是否有单一垂直、合作伙伴或账户主导支出 | 投资人可能把生态关注度误当作多元化客户经济性 | 索取头部垂直和头部地域收入组合,并配套活跃组织数 |
| 官方故障透明度和评价可见性若处理得当,可以建立信任 | 支持延迟、安全担忧和故障沟通混乱可能抬高流失率,或挡住受监管买方 | 在支持责任归属和事故应对姿态改善前,生产买方可能暂停扩张 | 索取支持 SLA 达成率、事故复盘节奏,以及与支持或可靠性事件相关的流失 |
各行将可能的先落地再扩张驱动因素,与让集中度和耐久性无法完整投资判断的具体公开数据缺口配对。
[CU032, CU034, CU035, CU036, CU037, CU040]6.4 客户结论与证据缺口
客户结论积极但有条件。OpenRouter 显然有真实需求,而且需求并非假设:独立媒体、公司自身披露、Sacra 早期估算和第三方开发者工具文档,都指向一个已经成为多模型 AI 用量的重要路由层的平台。今天证据最充分的客群,是看重快速模型访问和供应商可移植性的开发者、初创产品团队和 编程 agent 生态。资料最缺的是对高价值、耐久企业客户的直接证明。公开证据显示企业导向控制已经存在,公司也在尝试上行,但没有揭示有多少付费组织完成了这种迁移、收入集中度如何,或续约是否支撑基础设施叙事。换句话说,直接客户证明比融资和用量信号更薄。这不是本章的致命缺陷,但是真正的尽调缺口。下一步尽调应是向管理层要数据,而不是再做一次网页搜索:按支出带划分的活跃组织、直接与渠道组合、头部客户集中度、合同原型和 cohort 留存,才是把强需求故事变成可完整承销客户质量故事的缺失拼图。[CU017, CU018, CU029, CU030, CU035, CU036]
6.5 证据要点
07风险
7.1 整体风险排序与监管 / 法律包袱
OpenRouter 的剩余风险画像很特殊:它既是控制层,也成了新的集中点。产品承诺是靠路由带来更高可用性、更低成本和更好合规,但同一套架构也让模型供应商条款、路由政策、隐私设置、日志选择和客户支出治理都汇到 OpenRouter 这里。法律与合规风险因此不只是文书问题;一旦路由默认值、文档或策略开关被误解,错误会同时传导到许多供应商和许多客户。 眼下最直接的监管问题,不是 OpenRouter 本身显然属于被禁止的 AI 实践,而是它越来越像企业和代理式工作负载的底层管道,而这些工作负载正落入不断演进的 AI 治理义务。European Commission 的 AI Act 指引称,GPAI 义务已经生效,透明度义务将在 2026 年 8 月到来;高风险用例还需要日志、文档、人工监督和网络安全控制。OpenRouter 已经发布 EU 区域内路由、零数据留存、人工监督模式和数据处理条款等材料,但其中不少保护依赖配置,或只向企业客户开放,并非通用默认项。 这种选择加入式结构很重要。OpenRouter 自己的隐私和供应商日志材料说,公司不控制上游 LLM 如何处理输入和输出;部分供应商可能训练提示词,除非客户关闭;即便是不训练的端点,也可能出于滥用防控或法律原因留存数据。DPA 和隐私政策给出了通过 SCCs 和审计权的合同路径,但合同保护并不能消除执行风险。结果是,最大的法律 / 监管风险并不是已经出现的某一起诉讼,而是受监管流量可能以企业买方事后认为不合规或记录不足的方式被路由、留存、记录或解释。[CR001, CR002, CR003, CR004, CR005, CR007]
| 风险 | 公开证据 | 可能性 | 影响 | 缓解成熟度 | 剩余敞口 | 投资含义 | 尽调问题 |
|---|---|---|---|---|---|---|---|
| 受监管或欧盟流量被路由到获批提供商或司法辖区之外 | 欧盟专属路由和部分更严格控制仅面向企业或需主动启用;默认路由优先扩大覆盖,而不是采用最严格隐私姿态 | 中 | 高 | 中 | 高 | 可能卡住受监管工作负载扩张,或迫使团队为企业做定制打包 | 要求提供 eu.openrouter.ai 使用证明、提供商允许名单默认配置,以及无合规提供商场景下的例外处理 |
| 提供商留存或训练设置偏离客户预期 | 提供商日志说明,部分提供商若未关闭设置,可能把数据用于训练;即使有些端点不训练,也会因滥用防控或法律原因留存数据 | 高 | 高 | 中 | 高 | 会推高隐私敏感账户的流失和法务审查摩擦 | 核查提供商政策矩阵、账户默认值,以及有多少客户主动执行 data_collection deny 或 ZDR |
| AI Act / GPAI 合规负担拖慢企业采用 | 欧盟指引称,透明度义务将在 2026 年 8 月生效;高风险用途需要日志、文档、监督和网络安全措施 | 中 | 中 | 中 | 中 | 拉长企业销售周期,并抬高受监管用例的支持成本 | 询问哪些客户细分会触发 AI Act 审查,以及默认交付哪些合规材料 |
| 模型提供商条款导致访问暂停,或改变输出权利假设 | OpenRouter 条款把输出权利绑定到提供商模型条款;如果提供商要求处理,也允许暂停访问 | 高 | 中 | 低 | 中 | 模型政策冲击可能打断客户工作流、损害信任,且根因不由 OpenRouter 控制 | 要求提供历史模型下架、政策变更和客户沟通预案 |
| 跨境传输文档在尽调中受到质疑 | 隐私政策和 DPA 依赖 SCC 式传输机制,但买方仍需验证执行情况和子处理方 | 中 | 中 | 中 | 中 | 可能拖慢大型企业的安全审查或采购成交 | 要求提供 DPA 附表、子处理方明细,以及 NDA 下可提供的审计报告 |
| 输出或训练相关版权与 IP 责任仍未定 | Copyright Office 仍在梳理生成式 AI 训练、授权和责任问题 | 中 | 中 | 低 | 中 | 可能限制高价值企业用例,或要求更强的赔偿条款 | 要求说明当前赔偿立场,以及企业客户在输出使用和训练主张上的红线 |
可能性、影响、缓解成熟度和剩余暴露,是作者基于所引一手 / 法律来源及公司文档给出的定性评级;各行按剩余严重性排序,而不只按法律确定性排序。
[CR001, CR002, CR003, CR004, CR005, CR007]按残余风险呈现六项最可能削弱 OpenRouter 投资逻辑的风险敞口。
各单元格综合引用的产品文档、事故披露和监管指引;这些是分析师对公开可见缓释措施后的残余风险判断,不是公司发布的评分。
[CR023, CR024, CR028, CR032, CR033, CR037]7.2 运营、安全与可靠性风险
运营上,OpenRouter 最严重的风险是自己控制平面失效。2026 年 2 月 17 日和 19 日的复盘显示,第三方缓存依赖故障拖累了所有 API 入口,先是 500 错误,随后缓存失效把过多查询推向数据库,又表现成误导性的 401 认证错误。公司自己的复盘称,第二次宕机证明根因是缓存层,而不只是并发拒绝服务事件。这个信号在结构上很关键,因为 OpenRouter 的全部价值主张,就是让 AI 访问比接任何单一直接供应商更可靠。 更窄的状态事件进一步说明,问题不止一个根因。OpenRouter 披露过 Clerk 认证事件,影响登录和账户访问;Amazon Bedrock 宕机并传导到平台;另有一次覆盖聊天和生成 API 的 401 错误事件;还有一次近五小时日志事件,延迟了预算和账单事件。与此同时,OpenRouter 的路由文档说,供应商故障切换默认开启,但模型回退需要主动启用;出于合规原因限制供应商,会缩小可回退集合。换句话说,吸引受监管买方的同一套控制,也可能在上游系统失效时减少恢复路径。 公司确实有可信的缓释手段:熔断器、回退缓存、提示词注入阻断、敏感信息脱敏和支出上限。但多项防护是可选项,一些敏感信息检测仍像测试版,超时时可能放行请求。因此,投资人应把可靠性视为主风险:如果 OpenRouter 自身事故反复出现,平台就会从韧性层变成新的单点故障。[CR017, CR018, CR019, CR020, CR021, CR022]
| 失效模式 | 公开证据 | 可能性 | 影响 | 缓解成熟度 | 剩余暴露 | 投资影响 | 待澄清缺口 |
|---|---|---|---|---|---|---|---|
| OpenRouter 控制平面宕机会波及所有 API 入口 | 2026 年 2 月 17 日和 19 日的宕机源于第三方缓存故障,随后扩大为大范围 500/401 错误 | 中 | 高 | 中 | 高 | 若再次发生,公司的核心可靠性论点会受损 | 未披露公开可用性、SLO 历史和整改后事故率 |
| 计费 / 日志积压扭曲支出治理 | 状态页称,请求日志延迟也让预算和计费事件延迟了近五小时 | 中 | 高 | 中 | 高 | 计费争议会同时冲击信任、毛利率和支持负载 | 未公开 credits 调整或争议频率指标 |
| 认证或上游提供商事故削弱客户体验 | Clerk 登录降级和 Amazon Bedrock 宕机都传导成 OpenRouter 状态事故 | 中 | 中 | 中 | 中 | 即便推理不中断,依赖方事故也会增加支持负担 | 未公开认证和关键依赖的冗余或故障切换 SLA 图谱 |
| 合规受限路由压缩故障切换余量 | 提供商选择和主权路由文档显示,only/order/data 过滤器会缩小合格提供商集合 | 高 | 中 | 中 | 中 | 受监管客户可能用可用性换控制权,从而降低路由层价值 | 已在严格允许名单下运行的流量占比未知 |
| 提示注入或敏感数据泄露绕过控制 | Guardrails 可在移交给提供商前阻断 / 脱敏,但部分 NLP 检测器仍处 beta,超时可能放行请求 | 中 | 中 | 中 | 中 | 这类可见故障会伤害智能体和企业买方信任 | 未公开 guardrails 的误报 / 漏报性能数据 |
| 可观测性功能扩大提示词披露面 | I/O logging 和 Broadcast 可存储或导出完整提示词、输出、提供商名称和成本轨迹 | 中 | 中 | 中 | 中 | 配置错误可能增加隐私事故或客户顾虑 | logging / Broadcast 功能的默认状态和客户采用率未知 |
运营评级基于 2026 年直接事故证据和已记录的 guardrail 行为。若控制项为可选、仅企业版提供,或缺少公开性能指标,剩余暴露仍偏高。
[CR015, CR016, CR017, CR018, CR019, CR020]控制平面、隐私和计费故障如何传导为流失、利润率和估值压力。
边展示引用的事故、路由和计费来源描述或暗示的可能经济传导路径;这里不声称任何边有权重。
[CR029, CR030, CR031, CR032, CR033, CR037]7.3 伙伴依赖、脱媒与经济压缩
OpenRouter 依赖一整层交易对手方,而这些对象同时是供应商、政策制定者,有些情况下还是直接替代品。上游模型供应商仍决定许多客户最在意的硬约束:谁能处理数据、提示词是否可被留存或用于训练、适用哪些速率限制或配额,以及哪些条款约束输出所有权或服务暂停。OpenRouter 可以绕开一部分故障,但无法消除对供应商政策的依赖。其条款称,如果客户违反供应商条款,或供应商要求采取行动,模型访问可能被暂停;路由文档也写明,only/order/ignore 和 data_collection 过滤器等路由选择,是用韧性换控制。 经济风险还被放大了,因为大型平台现在卖的是原生路由,而不只是原始模型端点。Amazon Bedrock 提供跨模型家族的提示词路由;Microsoft Foundry 的模型路由器已经能在单一部署内按质量、成本和延迟优化。这让路由越来越像一个功能,而不是独立产品。再叠加 OpenAI 和 Google 的配额系统,即便上游约束在别处,OpenRouter 也可能成为被责怪速度慢、被限流或成本波动的一方。 认证、计费和支付通道也有更软但真实的伙伴依赖。状态页显示,Clerk 和 Bedrock 事件会在 OpenRouter 层浮现;条款页则把 Stripe 和 Coinbase 法律协议接入了计费栈的一部分。这些都不意味着模式已经坏掉;它意味着投资人应假设利润率压力、支持摩擦和伙伴政策冲击是商业模式的常态,除非 OpenRouter 证明即便存在直接路由替代品,也能让买方保持忠诚。[CR002, CR003, CR011, CR021, CR028, CR029]
| 依赖项 | 交易对手角色 | 失效情景 | 可能性 | 影响 | 缓解成熟度 | 剩余暴露 | 投资影响 |
|---|---|---|---|---|---|---|---|
| 上游前沿模型提供商 | OpenAI / Anthropic / Google 等模型托管方提供容量、数据政策条款和可用性 | 提供商政策变化、速率限制或宕机会移除有吸引力的路由路径,或迫使重新定价 | 高 | 高 | 中 | 高 | 提供商冲击并非 OpenRouter 能完全控制,但责任可能落到 OpenRouter 头上 |
| 超大云厂商原生路由器 | AWS Bedrock 和 Microsoft Foundry 已内置跨模型路由 | 路由变成平台捆绑功能,挤压 OpenRouter 抽成率和差异化 | 高 | 高 | 低 | 高 | 毛利率能否守住成为核心投资判断问题 |
| 认证提供商 | Clerk 把住登录和账户访问入口 | 即使 API 仍可用,认证降级也会阻断控制台访问并增加支持负载 | 中 | 中 | 中 | 中 | 运营摩擦会侵蚀开发者信任 |
| 缓存 / 内部控制平面依赖 | 第三方缓存承担面向数据库的快速 API key 查询中介 | 依赖故障会在所有 API 入口重现 2026 年 2 月故障模式 | 中 | 高 | 中 | 高 | 控制平面事故若重演,会直接削弱公司的可靠性主张 |
| 支付与 credits 通道 | Stripe 和 Coinbase 条款支撑部分 credits / 支付功能 | 条款变化、暂停服务或结算流程降级会扰乱 credits、开户或争议处理 | 中 | 中 | 低 | 中 | 在核心变现界面上增加本可避免的流失和支持摩擦 |
| 监管相关方 | EU AI / GDPR 监管机构决定哪些路由和日志姿态能卖给企业 | 新指引或审查会迫使默认值更严格、文档更多,或实施区域隔离 | 中 | 中 | 中 | 中 | 如果销售上探高端市场,合规成本可能比总量增长更快 |
集中度只能部分定性判断,因为 OpenRouter 未公开按提供商、支付通道或客户细分拆分的流量或收入;这种不透明本身就是风险因素。
[CR002, CR003, CR011, CR021, CR028, CR029]| 角色 / 经济暴露 | 当前缺口或压力 | 可能性 | 影响 | 缓解成熟度 | 剩余暴露 | 尽调路径 | 投资影响 |
|---|---|---|---|---|---|---|---|
| 抽成率韧性 | 超大云厂商路由器和提供商直连配额,会让路由逐步功能化 | 高 | 高 | 低 | 高 | 要求提供毛利率桥、返利经济性和赢单 / 输单定价数据 | 单靠量增可能掩盖单位经济性恶化 |
| 提供商集中度透明度 | 公开材料未披露按提供商划分的流量、支出或 GMV 集中度 | 高 | 高 | 低 | 高 | 要求提供头部提供商流量 / 支出拆分和集中度 guardrails | 单一提供商集中度飙升,可能让路由多元化变成幻象 |
| 企业合规打包 | 许多最强控制项是 opt-in 或仅企业版提供,而非对所有用户默认开启 | 中 | 中 | 中 | 中 | 询问企业账户中有多少默认使用 EU 路由、guardrails 和 ZDR | 如果买方需要定制配置帮助,销售效率可能走弱 |
| 支持与信任运营 | 计费边缘情况、误导性 401 和账户访问事故会增加客户支持负担 | 中 | 中 | 中 | 中 | 要求提供支持工单量、credits 调整率和重大事故复盘 | 即便量在增长,支持强度也会压缩利润率 |
| 安全 / 审计披露准备度 | 已审阅公开材料未提供完整、可直接用于尽调的认证和子处理方资料包 | 中 | 中 | 低 | 中 | 要求在 NDA 下提供 SOC2 / ISO 材料、渗透测试节奏和子处理方清单 | 披露缺失会拖慢采购,并削弱受监管工作负载转化 |
| 法律暴露准备度 | 本章审阅的直接材料中未找到公开诉讼案卷或执法资料包 | 低 | 中 | 低 | 中 | 由律师牵头,按法人实体和创始人筛查诉讼、制裁和隐私执法 | 意外法律问题对投资论点的重定价,可能高于当前文档暗示的程度 |
本表有意把执行风险和经济模型风险放在一起,因为 OpenRouter 的财务暴露与产品可靠性、提供商集中度和披露准备度密不可分。
[CR022, CR023, CR037, CR038, CR039, CR040]可能约束 OpenRouter 可用性、合规性或利润率的关键交易对手与政策制定者。
边表示运营或商业依赖,不代表正式排他。多个节点既是供应商也是替代品,战略风险因此更高。
[CR002, CR003, CR021, CR035, CR036, CR039]7.4 缓释措施、监控指标、推翻投资论点的触发点与尽调问题
OpenRouter 并没有无视这些风险;公开产品界面显示,公司在认真把缓释能力产品化。客户可以强制零数据留存路由、限制供应商、设置每个密钥和每个工作区的支出上限、让受监管工作负载走仅限 EU 的端点,并在流量到达供应商前阻断或脱敏提示词注入模式和敏感输入。宕机复盘也记录了具体修复动作,例如熔断器、更好的错误码和回退缓存。对承销判断来说,这些是有意义的正面因素,因为它们显示公司在把风险教训转成产品功能,而不是只靠流程兜底。 关键投资问题不是有没有缓释措施,而是这些措施是否足够默认、采用率是否足够高、是否足够可被独立审计,从而改变企业采购行为。多项控制只限企业版,一些需要主动启用,还有一些依赖客户纪律来正确配置。公开材料也留下重要盲点:按供应商的集中度、公开正常运行时间 / SLO 历史、独立安全证明,以及一套完整的尽调可用披露包,从外部看仍不完整。 因此,监控集合要保持具体。OpenRouter 自身 Sev-1 事故反复出现;证据显示受监管流量无法留在获批司法辖区或供应商内;账单抵扣争议持续;或超大云厂商原生路由器带来的赢单 / 输单压力上升,都会伤到投资论点。在承销关键任务或受监管采用前,投资人应在 NDA 下索取供应商集中度、SLA / SLO 数据、分处理方和认证细节,以及历史账单抵扣调整 / 争议指标。[CR013, CR017, CR018, CR019, CR020, CR021]
| 风险 | 可监控触发项 | 阈值 / 事件 | 行动含义 | 尽调要求 |
|---|---|---|---|---|
| 路由层可靠性 | OpenRouter 自身引发的重大事故 | 2026 年 2 月整改后,单季度发生超过一次由平台造成的重大宕机 | 在根因趋势被证明修复前,将可靠性视为投资论点已破 | 要求提供月度 Sev-1/Sev-2 日志、事后复盘和可用性 / SLO 历史 |
| 受监管数据合规 | 证据显示合规路由必须失败放行,或离开获批司法辖区 / 提供商 | 只要无法证明目标工作负载仅在 EU 处理、ZDR 或 data_collection-deny 行为 | 将使用限制在非敏感工作负载,并下调企业端上行空间 | 要求演示架构、日志和受监管流量样本审计证据 |
| 价格 / 抽成率压缩 | AWS、Microsoft 或模型厂商扩张原生路由,同时出现降价或返利压力 | 路由价格持续让步,却没有留存或增购收益抵消 | 下调利润率假设,并要求更强留存证明 | 要求提供 cohort 留存、混合抽成率,以及头部赢单 / 输单定价案例 |
| 计费完整性 | credits 争议或失败请求误收费 | 反复有投诉称,429 / 部分输出路径仍消耗 credits,且没有快速补救 | 上调支持成本假设,并对毛利率做压力测试 | 要求提供争议量、平均 credits 调整延迟和计费 QA 控制 |
| 提供商集中 | 流量或支出过度依赖一个前沿模型家族 | 任何单一提供商占流量或支出 >50%,且没有合同化兜底承诺 | 施加集中度折价,并质疑路由防御性 | 要求提供月度提供商组合仪表盘和合同灵活性摘要 |
| 合规披露准备度 | 尽调中缺少独立安全 / 隐私材料 | 无法在 NDA 下提供当前认证、渗透测试或子处理方文档 | 暂停受监管或关键任务场景的投资判断 | 要求提供完整信任资料包、审计报告和整改跟踪表 |
| 支付 / 账户依赖 | Stripe、Coinbase 或认证提供商政策 / 服务变化影响开户或 credits | 只要核心外部服务引发暂停、结算中断或长时间登录宕机 | 上调流失假设,并重新评估结账摩擦 | 要求提供依赖图谱、兜底方案,以及计费 / 认证界面的事故响应 |
| 法律 / 执法态势 | 出现新的隐私、AI 治理或 IP 挑战 | 任何与路由、留存或输出相关的正式监管询问、禁令或重大诉讼 | 在暴露边界确定前,暂停投资或要求重新定价 | 交割前由律师完成诉讼、隐私、制裁和 AI 治理尽调 |
这些门槛有意设计成可监控:它们帮助投资者区分 OpenRouter 的剩余风险只是运营噪音,还是投资论点应被暂停、重定价或判定破裂。
[CR020, CR021, CR022, CR031, CR037, CR043]7.5 图表证据
08估值
8.1 融资背景与价格支撑
OpenRouter 确实赢得了投资人注意。公司在 2026 年 5 月宣布 $113M Series B,独立报道把该轮对应到约 $1.3B 投后估值。官方材料也显示真实使用势头:每周 25T token、超过 8M 用户、接入 400+ 模型。这足以支撑严肃尽调,也解释了成长型投资人为什么想获得敞口。但它还不足以承销价格。公开备案轨迹仍然稀薄、滞后,并且围绕 2025 年 Form D,而不是当前经营报表。SEC 材料没有披露当前收入、烧钱、现金、客户集中度或优先股经济条款;已审阅的公开记录也没有说明新一轮是否包括老股转让、参与型优先权或期权池扩张。实际含义是,市场在给强战略相关性和需求动能定价,而公开证据仍让核心价格支撑问题悬而未决。[CV001, CV002, CV004, CV005, CV007, CV008]
| 建议 | 置信度 | 风险评级 | 估值立场 | 决策含义 | 目标持有期 / 退出逻辑 |
|---|---|---|---|---|---|
| 继续研究 | 中 | 高 | 偏高 | 未拿到私有经济数据,或入场价格没有明显改善前,不要按当前名义价格投资测算。 | 若持有 4-6 年,投资人仍应以至少约 3x gross 为目标;这意味着未来稀释前退出价值需走到 >$3.9B。 |
这是一个价格敏感判断:公司质地可能有吸引力,但公开材料对估值的支撑仍不完整。
[CV002, CV017, CV041, CV042, CV043]| 维度 | 投资论点 | 反论点 | 什么会改变判断 |
|---|---|---|---|
| 市场 | 智能体系统需要跨多个提供商处理成本、信任和模型选择逻辑,模型路由因此正在长成真实品类。 | 路由也可能一直只是大型云或开源栈里的功能,而不是可持久独立变现的利润池。 | 证明企业买方会持续为路由和治理付费,而不是把它当作商品化胶水。 |
| 产品 | OpenRouter 以中立控制平面切入:多提供商访问、故障切换和优化都实用。 | 中立路由逻辑比独家算力或自有模型更容易复制。 | 证明专有路由功能带来溢价赢单率和可量化付费意愿。 |
| 客户 | 8M+ 用户和每周 25T tokens 说明漏斗顶端拉力强,采用真实存在。 | 公开来源未拆分付费账户、企业客户标识或收入留存;使用量可能高估变现深度。 | 提供付费客户数量、cohort 留存、企业客户占比和头部客户集中度。 |
| 财务 | 抽成率机制、平台费和 BYOK 变现提供多条收入杠杆。 | 唯一公开收入锚点已经陈旧,且相对当前估值很小;利润率质量仍不透明。 | 提供当前收入、扣除提供商分成后的毛利率,以及现金消耗 / runway。 |
| 竞争 | 多模型成为结构性常态,客户又在意可移植性,OpenRouter 因此受益。 | 超大规模云厂商、Fireworks、Together 和自托管路由器可能压低价格,或拿走最有价值的工作负载。 | 证明更大的对手没有侵蚀企业交易中的附加销售、定价或续约质量。 |
| 风险 | 2026 年一轮强劲融资降低了短期偿付压力,也为商业化和可靠性建设提供资金。 | 宕机、隐性成本投诉和未披露的优先条款,都可能伤害转化、留存或未来融资议价能力。 | 披露 SLA 表现、支持成本结构,以及最新一轮的优先股条款栈。 |
各行按影响估值的维度配对最强正反论据,而不是按泛泛的公司质量排序。
[CV004, CV006, CV014, CV015, CV016, CV032]建议基于一条简单链条:品类相关性和使用证明确实存在,但经济性不透明和价格支撑缺口仍主导承销判断。
[CV004, CV017, CV032, CV034, CV041, CV043]8.2 可比基准与入场纪律
最好的公开基准并不是完美同业,但仍可用来框定纪律。Cloudflare 和 Datadog 展示了当投资人能看到耐久增长和披露财务时,受企业信任的软件控制平面可以拿到怎样的溢价。Snowflake、CoreWeave、GitLab、Fastly 和 Akamai 则说明,当业务看起来更偏基础设施、更成熟或结构性利润率更低时,倍数会跌多远。放在这条光谱上,OpenRouter 显然有战略相关性,但其公开披露更接近不透明一端,而不是高溢价软件一端。私有市场参照也讲同一个故事。Together AI 和 Fireworks 在 2026 年都融了大轮,但 Fireworks 还披露了 $4B 估值和超过 $280M 的年化收入,这种披露水平和规模是 OpenRouter 公开档案无法匹配的。因此入场纪律很重要:如果没有私下证据证明 OpenRouter 已把使用量转化为收入质量,当前头部价格就是让投资人在关键经济指标可见前先为上行付费。[CV018, CV019, CV020, CV021, CV022, CV023]
| 可比公司 | 指标 | 倍数 / 估值 / 状态 | 参考意义 | 局限 |
|---|---|---|---|---|
| Cloudflare | 公开市值 / TTM 收入 | 约 40.9x 收入 | 显示高端、受企业信任的控制平面能拿到怎样的估值。 | Cloudflare 是范围更广的安全与网络软件公司,披露更充分,护城河也完全不同。 |
| Datadog | 公开市值 / TTM 收入 | 约 25.6x 收入 | 可作为可观测性和平台软件质量的参考基准。 | 作为上市软件平台,Datadog 更多元、更成熟。 |
| Snowflake | 公开市值 / TTM 收入 | 约 19.3x 收入 | 可参照仍享受溢价的高增长数据基础设施。 | 不是路由业务;经济性和产品范围不同。 |
| CoreWeave | 公开市值 / TTM 收入 | 约 9.3x 收入 | 当前最接近的公开 AI 基础设施热度基准。 | 拥有算力和训练基础设施,因此不是中立控制平面。 |
| GitLab | 公开市值 / TTM 收入 | 约 5.7x 收入 | 作为开发者基础设施,可参考更常态化的软件倍数。 | 并非专门面向 AI 路由,且估值已按上市成熟度定价。 |
| Akamai / Fastly | 公开市值 / TTM 收入 | 约 3.9x / 约 4.5x 收入 | 用来校验低倍数边缘与交付基础设施的下行情形。 | 它们是成熟 CDN / 边缘厂商,不是模型路由平台。 |
| Together AI | 私有公司官方融资状态 | 2026 年宣布 $800M Series C 轮;所审阅官方来源未显示估值。 | 证实 2026 年投资人仍有胃口投 AI 基础设施同类公司。 | 官方来源在所审阅页面未披露估值或收入。 |
| Fireworks AI | 私有公司官方融资状态 | 2026 年估值 $4B,年化收入 >$280M,且此前有 $552M Series B 轮。 | 是本组中信息量最大的已披露私有 AI 基础设施参照。 | Fireworks 拥有更多基础设施栈,不是只做中立路由的业务。 |
样本按业务模式相关性和截至 2026 年 7 月可用估值证据筛选;私有路由市场在公开来源中披露并不完整。
[CV018, CV019, CV020, CV021, CV022, CV023]相关公开可比公司的收入倍数,从低个位数基础设施公司到高溢价软件控制平面不等;多数证据相关公司集中在 Cloudflare 这个异常值之下。
数值是用 2026 年运行日抓取的 CompaniesMarketCap 页面,按市值除以 TTM 收入简单计算得出。
[CV018, CV019, CV020, CV021, CV022, CV023]8.3 情景框架与投资判断
情景测算解释了为什么建议不是买入。在 $1.3B 入场价下,OpenRouter 仅要接近 25x 收入倍数,就需要约 $52M 收入;若要按 18x 交易,需要超过 $72M;若要落入较低溢价基础设施公司常见的 9x 倍数,则需要约 $144M。公开档案里唯一接近收入的锚,是 Sacra 在 2025 年 5 月估算的 $5M 年化收入;如果没有实质变化,当前估值看起来极其拉伸。当然,实质变化可能已经发生;官方使用增长提示了这种可能。但证明它所需的证据是私有的,不是公开的。因此,牛市情景可信但未验证。基本情景是,OpenRouter 具有战略重要性、商业前景不错,但仍过于不透明,难以形成价格驱动的确信。熊市情景是,该业务的经济属性更接近路由经纪商,而不是高溢价软件平台。这个组合支持继续研究的建议、中等信心、高风险和估值拉伸判断。[CV011, CV012, CV013, CV014, CV015, CV016]
| 情景 | 收入 / 质量假设 | 倍数 / 参照 | 隐含估值区间($M) | 相对 $1.3B 的毛回报 | 概率信号 / 下行触发因素 |
|---|---|---|---|---|---|
| 乐观 | 当前收入运行率已超过 $75M,企业客户留存强,利润率更像高端基础设施软件,而不是经纪中介。 | 18x-25x 收入,更接近 Snowflake / Datadog / 高端 AI 基础设施区间。 | 1,350-2,750 | ~1.0x-2.1x | 需要私有数据证明,使用量增长已快速转成高质量经常性收入。 |
| 基准 | 收入已有规模,但经济性仍像经纪中介;企业客户占比在改善,利润率和留存证据还不完整。 | 12x-18x 收入,低于高端软件,高于成熟基础设施。 | 540-1,260 | ~0.4x-1.0x | 如果管理层能证明真实增长,但还拿不出高端软件经济性,这是最匹配的情景。 |
| 悲观 | 收入仍相对有限,抽佣率经济性偏薄,客户对价格或可靠性的异议拖慢转化。 | 6x-10x 收入,更接近低质量基础设施或经纪中介结果。 | 120-350 | ~0.1x-0.3x | 如果隐性成本摩擦、反复宕机或留存疲弱暴露出低护城河的路由中介属性,该情景更可能成立。 |
区间是由证据牵引的估值框架,不是在声称当前收入已知;公开证据没有披露足以在这些情景中做选择的 2026 年实际收入运行率。
[CV036, CV037, CV038, CV039, CV040, CV042]现值框架显示,只有乐观情景能让当前明面估值舒适落在合理区间内。
区间结合了假设收入带和公开可比倍数;它们是估值框架,不是公司披露的预测。
[CV036, CV037, CV038, CV039, CV040, CV042]紧凑的 IC 式评分卡,展示公开材料下目前最强和最弱的位置。
标签是综合引用来源得出的、有证据支撑的判断,不是外部评分卡。
[CV006, CV017, CV025, CV031, CV041, CV043]8.4 退出准备度、最终尽调问题与推翻投资论点的触发点
退出准备度不只看 OpenRouter 能否增长,还要看未来买方或公开市场能否从今天的位置承销耐久经济性。新投资人若按报道估值入场,通常会希望多年持有,并看到通向远高于 $3.9B 价值的可信路径,才能在后续稀释前拿到 3x 总回报。这条路径并非不可能,但它需要公开档案里仍缺失的证据:当前年化收入、扣除供应商分成后的毛利率、留存、集中度、支持经济性,以及确切优先权栈。因此,最终尽调问题很直接,也关乎投资成败。若管理层能提供这些项目,并且它们证实高溢价软件经济性,判断可以上调。若材料反而显示低毛利、弱留存、反复宕机或惩罚性优先权栈,投资论点会很快破裂。投资人应把这些数据请求和终止触发点视为核心估值工作,而不是外围确认性尽调。[CV041, CV042, CV044, CV045, CV046, CV047]
| 触发因素 | 阈值 / 事件 | 如何击穿论点 | 行动含义 |
|---|---|---|---|
| 收入质量低于预期 | 当前收入运行率或利润率结构更接近低抽佣率经纪中介经济性,而不是高端软件经济性。 | 支付高端控制平面倍数的理由被击穿。 | 退出,或大幅下调估值预期。 |
| 留存 / 集中度风险 | NRR、客户流失或头部客户集中度显示企业客户基础脆弱。 | 削弱路由层的耐久性,并压窄退出可选路径。 | 暂停投资测算,直到客户质量被证明。 |
| 可靠性问题反复 | 在企业信任建立前,又出现一组严重宕机或 SLA 失效,影响生产客户。 | 伤害投资人愿意付钱的核心价值主张:可靠性和路由质量。 | 立即下调护城河判断和可比估值区间。 |
| 优先股条款栈苛刻 | 最新股权结构条款包含投资人保护、参与权或二级出售组合,实质性压低普通股等价上行空间。 | 即使名义估值不变,实际回报潜力也会下降。 | 重新定价,或放弃交易。 |
| 竞争胜率坍塌 | 超大规模云厂商、Fireworks 或自托管路由器持续赢走最高价值工作负载。 | 说明该品类仍会多宿主,OpenRouter 无法守住高端抽佣率。 | 将判断从平台溢价下调为商品化路由工具。 |
触发因素都与估值直接相关:每一项都会改变倍数或稀释计算,而不只是改变定性叙事。
[CV041, CV042, CV044, CV046, CV047]| 主题 | 缺失证据 | 重要性 | 责任方或尽调路径 |
|---|---|---|---|
| 当前收入 / ARR | 最新月度收入和年化收入,按自助、企业、路由量和 BYOK 变现拆分。 | 用来检验 $1.3B 价格落在合理可比区间内还是区间外。 | 向 CFO 或财务资料室索取。 |
| 扣除模型供应商分成后的毛利率 | 按流量类型拆分贡献毛利率,包括免费层补贴和支持负担。 | 决定业务更像高端软件、支付,还是薄利中介。 | 财务 + 运营工作流。 |
| 留存与扩张 | 毛留存、NRR、分群曲线和付费账户分层。 | 没有这些,使用量增长仍可能掩盖变现后的耐久性不足。 | 向收入运营 / 董事会材料索取。 |
| 股权结构和轮次条款 | Series B 轮优先权栈、参与权、反稀释、期权池调整,以及任何二级出售成分。 | 仅看名义估值不足以判断普通股等价上行空间。 | 法务尽调和股权结构审阅。 |
| 客户集中度 | 按收入排名的前 10 / 前 20 大客户,以及任何超大规模云厂商或渠道依赖。 | 集中度会压缩退出质量,并抬高续约风险。 | 客户分析 + 销售负责人。 |
| 可靠性和支持经济性 | SLA 表现、支持人员配置模式、宕机补救成本,以及企业支持附加率。 | 可靠性是产品承诺的一部分,守住它可能很贵。 | 工程 / 支持运营尽调。 |
每项索取都在补一个影响估值的关键缺口,而不是泛泛尽调愿望清单。
[CV008, CV010, CV041, CV042, CV045]免责声明
本报告是基于公开证据的尽调快照,不构成投资建议。重要财务、法律、技术和合同事实仍未公开;做出任何投资决定前,应直接向管理层核验,并查验一手文件。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | OpenRouter, Inc. is a Delaware corporation organized in 2023. | 高 | SO018, SO024 |
| CO002 | OpenRouter’s public legal address and principal place of business point to 169 Madison Avenue in New York, New York. | 高 | SO006, SO024 |
| CO003 | OpenRouter positions itself as the unified interface for LLMs and as an AI model exchange. | 高 | SO007, SO019 |
| CO004 | The product standardizes access to hundreds of models through one API while routing, failing over, and optimizing requests across providers. | 高 | SO007, SO011, SO019 |
| CO005 | OpenRouter’s enterprise pitch is unified access to 400+ models with one API, one bill, and a single contract. | 高 | SO003, SO004 |
| CO006 | OpenRouter publicly offers Free, Pay-as-you-go, and Enterprise pricing tiers. | 高 | SO003, SO004 |
| CO007 | Official and investor materials consistently frame OpenRouter as founded in 2023, with the about page saying it started in early 2023. | 高 | SO002, SO018, SO019 |
| CO008 | Alex Atallah is publicly identified as OpenRouter’s CEO and co-founder. | 高 | SO019, SO021, SO024 |
| CO009 | The SaaS News identifies Louis Vichy as a co-founder alongside Alex Atallah. | 中 | SO022 |
| CO010 | Reviewed public sources do not surface a broader named executive bench comparable to a mature enterprise infrastructure vendor. | 中 | SO002, SO019, SO031 |
| CO011 | The mirrored Form D lists Alexander Atallah as both an executive officer and a director. | 中 | SO024 |
| CO012 | The mirrored Form D lists Anjney Midha as a director. | 中 | SO024 |
| CO013 | Public materials reviewed do not disclose board committees, observer rights, or investor protective provisions. | 中 | SO006, SO024 |
| CO014 | OpenRouter announced a $113 million Series B on 2026-05-26/27 led by CapitalG. | 高 | SO011, SO019, SO020 |
| CO015 | The Series B syndicate included NVentures, ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, Databricks Ventures, AMP PBC, Pace Capital, Andreessen Horowitz, and Menlo Ventures. | 高 | SO011, SO019, SO021 |
| CO016 | Public reporting placed OpenRouter’s post-money valuation at approximately $1.3 billion after the Series B. | 高 | SO020, SO023 |
| CO017 | TechCrunch reported that OpenRouter had raised a $40 million Series A in June 2025 led by Andreessen Horowitz and Menlo Ventures with Sequoia participation. | 中 | SO020 |
| CO018 | Publicly reported rounds support at least $153 million of disclosed equity capital across the reported 2025 Series A and 2026 Series B. | 中 | SO011, SO020 |
| CO019 | The mirrored Form D showed a $47.606863 million equity offering with $29.606894 million sold to 30 investors and a first sale date of 2025-05-27. | 中 | SO024 |
| CO020 | The mirrored Form D declined to disclose OpenRouter’s revenue range and aggregate asset value. | 中 | SO024 |
| CO021 | Public sources do not reconcile whether the Form D proceeds fully overlap with the round later summarized in startup press, so the exact total raised remains under-disclosed. | 中 | SO020, SO024 |
| CO022 | Official funding materials said OpenRouter’s weekly volume grew from 5 trillion to 25 trillion tokens in six months. | 高 | SO011, SO019, SO020 |
| CO023 | Official funding materials said the platform was processing 100 trillion tokens per month, serving 8 million-plus global users, and spanning 400-plus models. | 高 | SO011, SO019, SO020 |
| CO024 | OpenRouter’s current official site counters later displayed 100 trillion monthly tokens, 10 million-plus global users, 70-plus providers, and 400-plus models. | 中 | SO001, SO002 |
| CO025 | OpenRouter’s quickstart docs describe an OpenAI-compatible endpoint that automatically handles fallbacks and selects the most cost-effective options. | 高 | SO007, SO014 |
| CO026 | Workspaces separate API keys, routing defaults, guardrails, observability, membership, and budgets inside one account. | 高 | SO008, SO012 |
| CO027 | Zero Data Retention can be enforced globally, by model group, by guardrail, or per request. | 高 | SO009, SO016 |
| CO028 | OpenRouter says prompt retention is opt-in while request metadata is stored to power reporting, rankings, and usage logs. | 中 | SO009, SO010 |
| CO029 | The May 2026 release spotlight tied the Series B to 100 trillion monthly tokens and to new workspace guardrails such as spend limits, provider allowlists, prompt-injection blocking, and PII redaction. | 高 | SO016, SO026 |
| CO030 | OpenRouter’s June 2026 image API launch added a dedicated API with unified access to 30-plus image models and per-endpoint pricing. | 高 | SO015, SO026 |
| CO031 | OpenRouter disclosed related outages on 2026-02-17 and 2026-02-19 that escalated from partial failures into near-total downtime for 38 and 35 minutes, respectively. | 中 | SO013 |
| CO032 | Post-incident remediations included circuit breakers, accurate 503 error codes for infrastructure failures, and fixes from the caching provider. | 中 | SO013 |
| CO033 | OpenRouter’s status site reported no incidents between May 2026 and July 2026 after the February outage write-up. | 中 | SO017 |
| CO034 | TrueFoundry’s 2026 review says production-stage concerns cluster around support delays, account security, and agentic workloads that can consume credits quickly. | 中 | SO025 |
| CO035 | TrueFoundry said Trustpilot showed a 1.7 out of 5 score across 41 reviews as of May 2026. | 中 | SO025 |
| CO036 | WisdomAI framed OpenRouter’s value as reducing vendor lock-in and letting teams route requests by task, cost, or compliance requirement. | 中 | SO027 |
| CO037 | OpenRouter’s GitHub organization shows active public Go, TypeScript, Python, agent, and example repositories updated through late June and early July 2026. | 中 | SO030 |
| CO038 | CapitalG’s portfolio page says OpenRouter sits between agents, applications, and the model ecosystem. | 高 | SO018, SO019 |
| CO039 | An archived April 2025 Tracxn profile described OpenRouter as Anguilla-based and backed by Soma Capital, conflicting with later New York and broader investor records. | 低 | SO029 |
| CO040 | The SaaS News described OpenRouter as a New York based AI infrastructure startup founded by Alex Atallah and Louis Vichy. | 中 | SO022 |
| CO041 | AI Market Watch says OpenRouter targets developers, startups, and companies building AI applications that want to avoid vendor lock-in. | 中 | SO028 |
| CO042 | AI Market Watch describes the product set as a unified REST API, model router, analytics dashboard, and standardized model format. | 中 | SO028 |
| CO043 | TechCrunch argued that OpenRouter’s growth indicates enterprises are moving toward multi-model AI stacks rather than a single all-powerful vendor. | 中 | SO020 |
| CO044 | Releasebot’s June 2026 rollup shows OpenRouter shipping MCP, image APIs, and other developer tooling in quick succession. | 中 | SO026 |
| CO045 | OpenRouter’s careers page says the company is a small remote-first team that routes billions of tokens every month and hires anywhere in the United States. | 中 | SO031 |
| CO046 | Reviewed public sources do not disclose a precise employee headcount. | 中 | SO024, SO031 |
| CO047 | The retained company-overview sources support adoption through user, token, and developer signals but do not identify named paying enterprise customers or case-study logos. | 中 | SO011, SO019, SO022 |
| CM001 | OpenRouter normalizes model and provider schemas around an OpenAI-like interface so developers can learn one API instead of many provider-specific ones. | 中 | SM001 |
| CM002 | OpenRouter exposes routing primitives such as a models array, fallback routing, and provider-selection controls, so its core product layer is orchestration rather than model hosting. | 中 | SM001, SM002 |
| CM003 | OpenRouter lets customers sort providers by price, throughput, or latency and can restrict routing to EU in-region processing or zero-data-retention endpoints for enterprise use cases. | 中 | SM002 |
| CM004 | OpenRouter prompt caching uses provider-sticky routing only when cache reads are cheaper than normal prompt pricing, embedding cost optimization into routing behavior. | 中 | SM003 |
| CM005 | IDC said that by 2028, 70% of top AI-driven enterprises will use advanced multi-tool architectures to dynamically manage model routing across diverse models. | 中 | SM004 |
| CM006 | IDC frames model routing value around performance, cost control, and governance or trust rather than around access to any single model. | 中 | SM004 |
| CM007 | Research and Markets sizes the broader AI gateway market at USD 3.66 billion in 2025 and USD 9.61 billion by 2032 with a 14.7% CAGR. | 中 | SM005 |
| CM008 | Intel Market Research sizes the narrower LLM middleware gateway market at USD 18.9 million in 2026 and USD 189 million by 2034 with a 49.6% CAGR. | 低 | SM006 |
| CM009 | The Business Research Company says the LLM gateway platform market grows from USD 3.34 billion in 2025 to USD 4.23 billion in 2026 and to USD 11.01 billion by 2030. | 中 | SM007 |
| CM010 | The same TBRC LLM gateway platform page also labels USD 11.01 billion as a 2035 revenue forecast in its report-attribute box, creating an internal date inconsistency. | 中 | SM007 |
| CM011 | The Business Research Company sizes the LLM gateways market at USD 2.18 billion in 2025, USD 2.76 billion in 2026, and USD 7.21 billion in 2030. | 中 | SM008 |
| CM012 | Intel Market Research sizes the AI API gateway market at USD 0.78 billion in 2025, USD 0.85 billion in 2026, and USD 2.12 billion by 2034. | 低 | SM009 |
| CM013 | The Research and Markets LLM router report treats the category as a broad hardware, software, and services market spanning data centers, cloud service providers, enterprises, and telecom. | 中 | SM010 |
| CM014 | Public estimates differ sharply because some sources measure narrow LLM middleware while others include broader AI gateways, gateway platforms, hardware, or enterprise control-plane services. | 中 | SM005, SM006, SM007, SM008, SM009, SM010 |
| CM015 | An OpenRouter-specific SAM should exclude most on-prem hardware, training infrastructure, and generic API-management spend because OpenRouter sells a managed developer-facing routing layer. | 中 | SM001, SM002, SM007, SM008, SM009, SM010 |
| CM016 | Included spend for OpenRouter-like services is unified API access, provider routing, fallback, caching, observability, and consolidated billing across third-party models. | 中 | SM001, SM002, SM003, SM011, SM014 |
| CM017 | Excluded spend includes model creation and training, GPU hardware, non-AI API programs, and end-application AI software revenue. | 中 | SM005, SM007, SM010, SM020 |
| CM018 | Status-quo substitutes for OpenRouter include direct provider APIs, cloud-native managed routers, AI-enabled API gateways, and self-hosted open-source routing stacks. | 中 | SM017, SM019, SM020, SM021, SM024, SM025 |
| CM019 | Cloudflare AI Gateway offers analytics, logging, caching, rate limiting, request retries, model fallback, and support for multiple providers through a lightweight managed layer. | 中 | SM011 |
| CM020 | Cloudflare makes core AI Gateway features free but charges a 5% fee on unified-billing credits and adds paid-plan request pricing above 10 million monthly requests. | 中 | SM012 |
| CM021 | Cloudflare caching is exact-match against provider, endpoint, model, auth header, and full request body, which helps repetitive traffic but leaves less room for semantic reuse. | 中 | SM013 |
| CM022 | Kong positions AI Gateway as a governance and connectivity layer with a provider-agnostic API, central credentials, dynamic routing, access control, and usage analytics. | 中 | SM014 |
| CM023 | Kong adds PII sanitization, prompt guarding, audit-ready observability, and other policy plugins that are designed for compliance-heavy production environments. | 中 | SM014, SM016 |
| CM024 | Kong AI Proxy standardizes multiple provider formats into an OpenAI-like interface and can also fulfill requests to self-hosted models. | 中 | SM015 |
| CM025 | Amazon Bedrock prompt routing gives customers a single serverless endpoint for quality-versus-cost routing inside a model family, but its preview scope is limited to selected models and it is optimized only for English prompts. | 中 | SM017 |
| CM026 | AWS and Kong argue that enterprise-scale AI needs centralized governance, token-based cost control, observability, and multi-model orchestration instead of many point-to-point provider accounts. | 中 | SM018 |
| CM027 | Microsoft Foundry model router is a single deployment with Balanced, Cost, and Quality modes, automatic failover, policy-aware model subsets, and prompt caching when the underlying models support it. | 中 | SM019 |
| CM028 | Azure API Management can expose multiple AI backends through one OpenAI-compatible endpoint and apply governance policies once across all routed models. | 中 | SM020 |
| CM029 | Azure AI gateway supports token quotas, semantic caching, load balancing, circuit breaking, monitoring, and audit-friendly token metrics across many applications. | 中 | SM020 |
| CM030 | LiteLLM supports weighted, latency-based, rate-limit-aware, least-busy, and lowest-cost routing across multiple provider deployments. | 中 | SM021 |
| CM031 | LiteLLM fallbacks can shift traffic across model groups after retries and can be configured for general errors, content-policy violations, or context-window failures. | 中 | SM022 |
| CM032 | TrueFoundry says enterprise buyers now evaluate gateways on governance, deployment model, security and RBAC, cost controls, and observability rather than on model breadth alone. | 低 | SM023 |
| CM033 | TrueFoundry characterizes OpenRouter as a strong fit for experimentation and model discovery but a weaker fit for governance-heavy and self-hosted enterprise production use cases. | 低 | SM024 |
| CM034 | Not Diamond describes OpenRouter as the de facto LLM marketplace with unified billing and strong failover, but says it has no self-hosted option and charges a 5.5% platform fee. | 低 | SM025 |
| CM035 | Maxim says well-executed multi-model routing can reduce token spend by 40% to 70% on mixed workloads while improving reliability through cross-provider failover. | 低 | SM026 |
| CM036 | Maxim says production buyers evaluate routing logic, performance overhead, provider coverage, failover, budgets, access control, observability, and deployment model. | 低 | SM026 |
| CM037 | Maxim’s enterprise-gateway comparison places OpenRouter at the easy-access end of the market, with less governance depth than self-hosted enterprise gateways. | 低 | SM027 |
| CM038 | Google positions Gemini Flash-Lite as its most cost-efficient, low-latency option for high-volume traffic while reserving more capable tiers for harder reasoning tasks. | 中 | SM028 |
| CM039 | A unified API lowers switching costs because teams can swap models or providers behind one abstraction instead of rewriting provider-specific integrations each time. | 中 | SM001, SM024, SM027 |
| CM040 | Switching costs are not eliminated because enterprise buyers may still need VPC deployment, internal key management, audit trails, policy enforcement, or data-residency guarantees that managed routers do not fully satisfy. | 中 | SM020, SM023, SM024, SM025, SM027 |
| CM041 | Developer-led teams and startups are the earliest natural users of managed routers, while platform engineering or AI infrastructure groups become more important as usage becomes shared infrastructure. | 中 | SM024, SM026, SM027 |
| CM042 | In larger organizations, finance or IT governance becomes a co-payer once token quotas, chargeback, and budget policies must be coordinated across many teams. | 中 | SM018, SM020, SM023 |
| CM043 | Common adoption triggers are provider sprawl, repeated API integration work, model price dispersion, outage or rate-limit risk, and the need for centralized analytics. | 中 | SM004, SM011, SM018, SM024, SM026 |
| CM044 | Governance and trust now matter because gateway vendors advertise zero-data-retention filters, EU routing, audit logs, managed identities, policy enforcement, and content safety as core features. | 中 | SM002, SM014, SM016, SM020 |
| CM045 | Inference economics and model-tier pricing make routing valuable because organizations do not want to send every prompt to the most expensive model when cheaper tiers are adequate. | 中 | SM017, SM026, SM028 |
| CM046 | Native clouds, API-gateway incumbents, and self-hosted routers cap third-party penetration because many buyers can satisfy basic routing, quota, or governance needs inside existing estates. | 中 | SM011, SM014, SM017, SM019, SM020, SM021 |
| CM047 | OpenRouter’s SaaS-only architecture and lighter governance depth constrain its reach in regulated or centrally governed enterprise environments. | 中 | SM024, SM025, SM027 |
| CM048 | OpenRouter’s relevant 2026 SAM is best treated as a low-hundreds-of-millions slice inside the broader gateway and control-plane market rather than as a whole-market headline. | 低 | SM005, SM006, SM007, SM008, SM009, SM010, SM024, SM027 |
| CM049 | A reasonable evidence-constrained 2026 SAM range is roughly USD 0.20 billion to USD 1.00 billion because it sits above narrow middleware estimates but below broader AI API gateway and gateway-platform categories while excluding self-hosted and hardware-heavy segments. | 低 | SM005, SM006, SM007, SM008, SM009, SM010 |
| CM050 | A tentative near-term SOM for independent managed routers is materially smaller than the constrained SAM because native clouds, enterprise AI gateways, and self-hosted LiteLLM-like stacks absorb much of the governance-heavy demand and contract budget. | 低 | SM017, SM019, SM020, SM021, SM024, SM027 |
| CP001 | OpenRouter routes requests to the best available providers for a model and load balances across top providers by default to maximize uptime. | 中 | SP022 |
| CP002 | OpenRouter's provider object can sort and filter providers, including quantization preferences. | 中 | SP022 |
| CP003 | OpenRouter model fallbacks automatically try other models when the primary providers are down, rate-limited, or blocked by moderation. | 中 | SP023 |
| CP004 | OpenRouter service_tier offers flex for lower cost and priority for faster service at higher cost. | 中 | SP024 |
| CP005 | OpenRouter documents a flex example that trades latency and availability for a 50% discount on select OpenAI requests. | 中 | SP024 |
| CP006 | OpenRouter Pareto Router selects coding models using a min_coding_score threshold instead of pinning one fixed model. | 中 | SP025 |
| CP007 | OpenRouter Fusion Router runs multiple models in parallel and uses a judge model to synthesize consensus, contradictions, and gaps. | 中 | SP026 |
| CP008 | Together AI markets an AI-native cloud with serverless inference, batch inference, dedicated model inference, dedicated containers, and GPU clusters. | 中 | SP001, SP002 |
| CP009 | Together AI homepage advertises a Series C announcement as of the 2026 run date. | 中 | SP001 |
| CP010 | Together serverless runs through a shared per-token API with no provisioning and no minimum cost. | 中 | SP002 |
| CP011 | Together says serverless is best for variable or bursty traffic, while steady or higher-rate workloads should move to dedicated endpoints. | 中 | SP002 |
| CP012 | Together says some batch workloads can be discounted up to 50% relative to real-time serverless. | 中 | SP002 |
| CP013 | Replicate markets an API to run and fine-tune models and deploy custom models. | 中 | SP003, SP005 |
| CP014 | Replicate homepage shows multi-million run counts on popular models, signaling broad usage. | 中 | SP003 |
| CP015 | Replicate pricing says some models bill by hardware-time and others by input/output tokens. | 中 | SP004 |
| CP016 | Fireworks says it processes 30T+ tokens per day and optimizes inference throughput and latency. | 中 | SP006 |
| CP017 | Fireworks bundles serverless inference, on-demand deployments, and training on one platform. | 中 | SP006, SP007, SP008 |
| CP018 | Fireworks pricing is per-token for serverless and per-GPU-second for on-demand deployments. | 中 | SP007 |
| CP019 | Martian says it builds model routers that dynamically select the optimal AI model for each query. | 中 | SP011 |
| CP020 | Martian says its routing optimizes performance, cost, uptime, and other business requirements. | 中 | SP011 |
| CP021 | Martian's Accenture partnership and Airlock compliance launch imply an enterprise and regulated-workload motion. | 中 | SP011 |
| CP022 | RouterBench argues that no single model is optimal across capability, latency, and cost tradeoffs. | 中 | SP010 |
| CP023 | LiteLLM markets an OpenAI-format gateway across 100+ LLMs. | 高 | SP012, SP013, SP015 |
| CP024 | LiteLLM includes fallbacks, spend tracking, budgets or rate limits, and routing or load balancing. | 中 | SP012, SP013, SP014 |
| CP025 | LiteLLM repo metadata shows roughly 52k stars and 9k+ forks as of 2026-07-01. | 中 | SP015 |
| CP026 | Amazon Bedrock says it serves more than 100,000 organizations worldwide and offers access to hundreds of foundation models. | 中 | SP016 |
| CP027 | Amazon Bedrock markets enterprise security, privacy, and compliance including ISO, SOC, GDPR, FedRAMP High, and HIPAA eligibility. | 中 | SP016 |
| CP028 | Amazon Bedrock prompt routing uses a single serverless endpoint to route between models within the same model family for quality and cost optimization. | 中 | SP017 |
| CP029 | Amazon Bedrock pricing varies by model and provider and says select batch inference is 50% lower than on-demand pricing. | 中 | SP018 |
| CP030 | Microsoft Foundry positions itself as a unified AI platform for models, agents, trustworthy AI, token controls, and governance. | 中 | SP019, SP021 |
| CP031 | Azure model router is a single deployment that routes prompts to eligible models by complexity or task while honoring access and data-zone boundaries. | 中 | SP020 |
| CP032 | Azure model router exposes Balanced, Cost, and Quality modes that trade off quality bands against cost. | 中 | SP020 |
| CP033 | OpenAI API pricing publishes per-1M-token prices and multiple service variants such as standard, batch, flex, and priority. | 中 | SP027 |
| CP034 | Anthropic pricing publishes token-metered API economics plus enterprise security and admin options such as SSO, SCIM, and HIPAA-ready offerings. | 中 | SP028 |
| CP035 | LlamaIndex exposes router modules as part of its developer framework. | 中 | SP029 |
| CP036 | LangChain emphasizes model portability and configurable routing or middleware in its agent harness. | 中 | SP030 |
| CP037 | OpenRouter's routing breadth is broader than Bedrock or Azure because it steers across top providers, while Bedrock stays within a model family and Azure routes within eligible deployments. | 高 | SP022, SP017, SP020 |
| CP038 | Inference clouds like Together and Fireworks compete by bundling routing-adjacent access with underlying compute, training, or dedicated infrastructure. | 中 | SP001, SP002, SP006, SP007 |
| CP039 | Replicate is closer to a model runtime or marketplace substitute than a neutral multi-provider router. | 中 | SP003, SP004, SP005 |
| CP040 | LiteLLM plus framework routers reduce switching costs because teams can standardize on OpenAI-style APIs or routing primitives without buying OpenRouter's managed control plane. | 中 | SP012, SP013, SP014, SP029, SP030 |
| CP041 | Hyperscalers have stronger distribution power because routing is sold inside broader cloud and AI platform contracts plus governance stacks. | 中 | SP016, SP019, SP021 |
| CP042 | OpenRouter's moat relies on software logic—provider selection, failover, cost or latency tiers, and specialized routers—rather than exclusive model supply. | 中 | SP022, SP023, SP024, SP025, SP026 |
| CP043 | Cloud incumbents currently hold the strongest trust and regulatory posture in this field. | 高 | SP016, SP019, SP020, SP021 |
| CP044 | Public pricing transparency is materially better for direct APIs, Replicate, Fireworks, and Together serverless than for Martian or enterprise cloud commitments. | 中 | SP002, SP004, SP007, SP021, SP027, SP028 |
| CP045 | Because no single model is optimal across cost, capability, and latency, buyers have reason to multi-home across providers, which supports the routing category but weakens lock-in for any one router. | 高 | SP010, SP022, SP024, SP029, SP030 |
| CI001 | OpenRouter announced a $113 million Series B led by CapitalG with participation from NVentures, ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, Databricks Ventures, AMP PBC, Pace Capital, Andreessen Horowitz, and Menlo Ventures. | 高 | SI001, SI002, SI003 |
| CI002 | OpenRouter said weekly volume grew from 5 trillion to 25 trillion tokens in the six months before the May 2026 Series B announcement. | 高 | SI001, SI002, SI003 |
| CI003 | OpenRouter said it serves more than 8 million developers or global users across 400-plus models. | 高 | SI001, SI002, SI003 |
| CI004 | OpenRouter describes itself as the routing layer between agents or applications and model providers, emphasizing routing, reliability, cost optimization, and compliance. | 高 | SI001, SI002, SI004 |
| CI005 | OpenRouter said it plans to use the Series B proceeds to scale infrastructure, deepen enterprise capabilities, and invest further in intelligent routing and governance. | 高 | SI001, SI002 |
| CI006 | OpenRouter’s SEC Form D materials identify the issuer as OpenRouter, Inc., incorporated in Delaware in 2023 with principal place of business in New York, and they show revenue declined to disclose. | 高 | SI013, SI014, SI015 |
| CI007 | The Form D filed on August 25, 2025 lists a first sale date of May 27, 2025, total offering amount of $47,606,863, total amount sold of $29,606,894, and 30 investors. | 高 | SI013, SI014, SI015 |
| CI008 | OpenRouter’s platform-fee announcement says non-crypto credit purchases are charged 5.5% of order value with a $0.80 minimum fee, while crypto purchases are charged 5.0% flat with no minimum. | 高 | SI016, SI021, SI024 |
| CI009 | OpenRouter’s BYOK announcement says the first 1 million BYOK requests per month are free and standard 5% BYOK fees apply beyond that threshold. | 高 | SI017, SI021, SI024 |
| CI010 | OpenRouter’s BYOK announcement says BYOK is supported for more than 60 inference providers and can either fail over to OpenRouter capacity or stay strictly on customer-owned keys. | 中 | SI017 |
| CI011 | OpenRouter’s service-tier documentation says requests can select flex or priority processing tiers and are billed at the rate of the actual tier that serves the request. | 中 | SI006 |
| CI012 | The same service-tier documentation gives an official example where flex can provide a 50% discount in exchange for higher latency and lower availability. | 中 | SI006 |
| CI013 | OpenRouter says it minimizes gateway latency through Cloudflare Workers, edge caching of user and API-key data, and optimized routing logic. | 中 | SI012 |
| CI014 | OpenRouter says low credit balances or API keys nearing their credit limit trigger extra database checks and more aggressive cache expiry, which can increase latency until new credits are added. | 中 | SI012 |
| CI015 | OpenRouter says failed initial model or provider completions can add latency because fallback tries another option. | 中 | SI012 |
| CI016 | The credits API tracks total credits purchased and used, indicating that OpenRouter maintains a wallet-style balance between customer prepayment and later consumption. | 中 | SI005 |
| CI017 | OpenRouter’s analytics and user-activity endpoints expose endpoint-level activity, metrics, dimensions, filters, and time ranges to management-key holders. | 中 | SI009, SI010 |
| CI018 | OpenRouter’s workspaces announcement says workspaces split keys, routing defaults, guardrails, BYOK, observability, and members while billing still rolls up at the account level. | 中 | SI020 |
| CI019 | The workspace budgets API documents that each workspace can have configured budgets. | 中 | SI008, SI020 |
| CI020 | Stripe Projects documentation says OpenRouter can be provisioned on a free plan by default and upgraded to pay-as-you-go per-token pricing with unified billing through Stripe. | 中 | SI007 |
| CI021 | OpenRouter’s free-tier update says the company is directly covering some free-model costs to keep popular models freely accessible. | 中 | SI018 |
| CI022 | OpenRouter’s February 2026 outage post says third-party cache failures caused 500 and 401 errors, with roughly 20% to 90% of API requests failing during 35- to 38-minute incidents. | 中 | SI019 |
| CI023 | OpenRouter’s outage post says the company deployed circuit breakers and a fallback-caching path after the February 2026 incidents. | 中 | SI019 |
| CI024 | TechCrunch reported that the May 2026 Series B implied roughly a $1.3 billion post-money valuation and compared it with an estimated $547 million post-money valuation after a $40 million Series A in June 2025. | 中 | SI003 |
| CI025 | Sacra estimated that OpenRouter reached $5 million of annualized revenue in May 2025 on $100 million of GMV, up from $1 million annualized revenue at the end of 2024. | 中 | SI023 |
| CI026 | Sacra said OpenRouter customers buy inference tokens across providers and that tokens are marked up roughly 5% while the product also provides a single dashboard for usage and billing. | 中 | SI023 |
| CI027 | TrueFoundry’s pricing review describes OpenRouter’s public plans as free tier plus pay-as-you-go, with 25-plus free models, 20 requests per minute on free models, and higher daily free-model limits once accounts hold at least $10 of credits. | 低 | SI021, SI026 |
| CI028 | TrueFoundry’s pricing review says enterprise terms are custom and oriented around SSO, SLA, and dedicated-support needs rather than pure self-serve usage. | 低 | SI021 |
| CI029 | TrueFoundry’s review summary says production complaints cluster around support delays, account security concerns, free-tier rate limits, and agentic workloads that can rapidly consume credits. | 低 | SI022 |
| CI030 | CostBench’s hidden-cost analysis says specialized providers can price some models far above cheaper alternatives on OpenRouter and that free models can be fragile under load. | 低 | SI025 |
| CI031 | ofox.ai’s pricing guide says the $0.80 minimum fee makes small top-ups materially more expensive in percentage terms and that failed routing retries can sometimes produce charges from both attempts. | 低 | SI024 |
| CI032 | CapitalG’s portfolio description says OpenRouter offers spend visibility, routing permissions, audit-friendly usage reporting, automated failover, and reduced vendor lock-in. | 中 | SI004 |
| CI033 | Public evidence supports at least four monetization layers: prepaid non-BYOK inference spend, platform fees on credit purchases, BYOK overage fees beyond 1 million requests, and higher-touch enterprise contracts or support. | 中 | SI016, SI017, SI020, SI021, SI023 |
| CI034 | The public record does not disclose revenue mix across pay-as-you-go, BYOK, and enterprise contracts. | 中 | SI001, SI016, SI017 |
| CI035 | Prepaid credits likely create favorable working capital but also gross-versus-net recognition ambiguity because cash can be collected before the underlying provider cost is incurred or the credit is consumed. | 中 | SI005, SI016, SI020, SI023 |
| CI036 | Margin drivers likely include platform-fee capture, BYOK overage fees, routing efficiency, free-model subsidy levels, and the support burden needed to keep enterprise workloads reliable. | 中 | SI016, SI017, SI018, SI019, SI023 |
| CI037 | OpenRouter appears structurally lighter on capex than model builders because its public materials emphasize routing software, caching, analytics, and failover rather than owned GPU fleets, even though reliability engineering still requires material operating spend. | 中 | SI001, SI002, SI012, SI019 |
| CI038 | The public GTM motion looks product-led at the top of funnel and enterprise-led on expansion, using free access, self-serve credits, Stripe provisioning, and later workspaces, budgets, and negotiated support. | 中 | SI007, SI017, SI020, SI021 |
| CI039 | The best public sales-efficiency proxies are rapid usage growth, 8 million-plus users, low integration friction, and management analytics, but CAC, payback, ACV, win rates, and net retention remain undisclosed. | 中 | SI001, SI003, SI007, SI009, SI010 |
| CI040 | Capital adequacy looks favorable in the near term after the $113 million Series B and the $1.3 billion valuation signal, but public sources do not disclose current cash balance, monthly burn, runway, debt, or contractual commitments. | 中 | SI001, SI002, SI003, SI014 |
| CI041 | The visible SEC footprint is too sparse for full underwriting because it is limited to a 2025 Form D and issuer metadata rather than GAAP statements or 2026 financing detail. | 高 | SI013, SI014, SI015 |
| CI042 | The main diligence blockers are ARR and revenue growth, gross-versus-net revenue policy, gross margin and provider COGS, burn and runway, enterprise contract mix, customer concentration, NRR or churn, and the economics of subsidized free models. | 中 | SI005, SI016, SI017, SI018, SI019, SI023 |
| CE001 | OpenRouter exposes a unified API that gives applications access to hundreds of models through a single endpoint while handling fallbacks across providers. | 高 | SE001, SE002, SE030 |
| CE002 | OpenRouter officially documents three integration paths: direct API calls, Client SDKs, and the higher-level Agent SDK. | 中 | SE001, SE019, SE020 |
| CE003 | The public product surface extends beyond chat completions into model discovery, rankings, routing plugins, server tools, structured outputs, observability, and workspace controls. | 中 | SE001, SE003, SE009, SE012, SE014, SE021, SE022 |
| CE004 | OpenRouter positions its endpoint as OpenAI-compatible so existing SDK and framework patterns can be reused instead of building new provider-specific adapters. | 高 | SE001, SE030, SE035 |
| CE005 | The models API exposes filters for output modality, pricing, context length, provider, author, architecture, benchmark ordering, distillability, ZDR, and region. | 中 | SE002, SE017 |
| CE006 | The models API can sort server-side by price, context, throughput, latency, popularity, and benchmark-derived ranking indexes. | 中 | SE017 |
| CE007 | Auto Router can pin both the selected model and provider across turns using either implicit conversation fingerprints or an explicit session_id, with stickiness expiring after five minutes of inactivity. | 高 | SE006, SE022 |
| CE008 | Auto Router exposes allowed_models filtering and a cost_quality_tradeoff control so teams can bound which models the router may select. | 高 | SE006, SE022 |
| CE009 | Auto Exacto runs automatically on tool-calling requests and requires no client-side configuration. | 中 | SE005 |
| CE010 | Auto Exacto reorders providers for tool-calling requests using tool-call success rates, throughput, and benchmark signals. | 中 | SE005 |
| CE011 | Auto Exacto measures request-level tool-call errors using invalid JSON, unknown tool names, and invalid arguments as explicit failure buckets. | 中 | SE005 |
| CE012 | Response caching happens at the OpenRouter layer before a request reaches any upstream provider. | 中 | SE007 |
| CE013 | When an identical successful request is served from cache, OpenRouter reports zero billable usage and returns cache status and TTL metadata. | 中 | SE007 |
| CE014 | OpenRouter excludes errors, rate-limit responses, partial results, and tool-calling requests from response caching. | 中 | SE007 |
| CE015 | Tool calling is standardized across providers and OpenRouter requires the tools schema to be present on both the initial request and the follow-up request that returns tool results. | 高 | SE008, SE030 |
| CE016 | Structured outputs on OpenRouter use JSON Schema validation to enforce machine-readable responses on compatible models. | 高 | SE012, SE030 |
| CE017 | Server tools are OpenRouter-operated capabilities that execute server-side when a model invokes them during a request. | 中 | SE009 |
| CE018 | The documented server-tools catalog includes web search, web fetch, image generation, workspace file operations, and panel-style multi-model analysis. | 中 | SE009 |
| CE019 | The web search server tool defaults to an auto engine that uses native provider search when available and otherwise falls back to Exa. | 中 | SE010 |
| CE020 | The web fetch server tool can use native provider fetch or alternate fetch engines and is also exposed through the Responses API. | 中 | SE011 |
| CE021 | Input and Output Logging is admin-gated, stores prompts and completions in isolated Google Cloud Storage, and uses AES-256 encryption at rest. | 中 | SE014 |
| CE022 | Logged prompt and completion content is retained for at least three months, can be deleted on request, and is skipped when traffic uses EU routing. | 中 | SE014 |
| CE023 | Account-level settings can exclude providers that may train on prompts, with separate controls for paid and free models. | 中 | SE015 |
| CE024 | Provider data-retention policies are exposed to users, but OpenRouter does not automatically enforce routing based on those retention requirements. | 中 | SE015 |
| CE025 | Enterprise EU routing uses the eu.openrouter.ai base URL to keep prompts and completions within the European Union. | 中 | SE015 |
| CE026 | The ZDR endpoint preview returns per-endpoint latency, throughput, uptime, pricing, supported parameters, and implicit-caching flags for eligible routes. | 中 | SE018 |
| CE027 | OpenRouter continuously monitors provider response times, error rates, and availability to inform routing decisions. | 高 | SE016, SE023 |
| CE028 | Provider failover is automatic and on by default within a chosen model, while model fallbacks require an explicit models array. | 高 | SE016, SE023 |
| CE029 | Provider outages and 429s are handled at the provider layer, while context-window and moderation failures require model-level fallback. | 中 | SE023 |
| CE030 | Zero completion insurance is enabled by default and waives charges for blank or failed zero-token responses. | 高 | SE013, SE023 |
| CE031 | The dedicated Image API provides one normalized request schema across more than 30 image models from eight providers. | 高 | SE021, SE022 |
| CE032 | Image endpoints expose typed capability descriptors, streaming support, granular pricing arrays, and provider-specific passthrough allowances. | 中 | SE021 |
| CE033 | OpenRouter’s May 2026 workspace guardrails added spend limits, model and provider allowlists, OWASP-derived prompt-injection blocking, and PII redaction. | 中 | SE022 |
| CE034 | OpenRouter’s May 2026 releases added speech-to-text and text-to-speech APIs with built-in provider failover and upstream error passthrough. | 中 | SE022 |
| CE035 | Model Fusion is available as an API plugin, server tool, and chatroom composer feature for parallel multi-model synthesis. | 中 | SE022 |
| CE036 | OpenRouter’s May 2026 enterprise controls included BYOK management APIs, per-provider ZDR controls, and session-id stickiness across turns. | 中 | SE022 |
| CE037 | OpenRouter maintains official GitHub, npm, and PyPI surfaces for its AI SDK provider, TypeScript SDK, Agent SDK, and Python SDK. | 中 | SE024, SE025, SE026, SE027, SE028, SE029 |
| CE038 | The TypeScript SDK is ESM-only and documents chat.send, pagination, file uploads, debug logging, and provider options such as zdr and sort. | 中 | SE025, SE028 |
| CE039 | The Agent SDK handles multi-turn loops, stop conditions, tool execution, and conversation state through callModel. | 中 | SE020, SE026 |
| CE040 | The AI SDK provider package supports embeddings, providerOptions and extraBody controls, usage accounting, response healing, and fine-grained tool streaming. | 中 | SE024, SE027, SE032 |
| CE041 | LangChain’s ChatOpenRouter integration documents tool calling, structured outputs, multimodal inputs, streaming, and reasoning metadata. | 中 | SE030 |
| CE042 | LiteLLM documents OpenRouter support for text, chat, vision, embeddings, image generation, and OpenRouter-specific route and transforms parameters. | 中 | SE031 |
| CE043 | OpenHands uses LiteLLM to call OpenRouter and accepts openrouter slash model strings for custom model selection. | 中 | SE033 |
| CE044 | Cline exposes OpenRouter as a configurable provider, supports custom base URLs, and documents a Gemini caching workaround. | 中 | SE034 |
| CE045 | Pydantic AI documents OpenAI-compatible providers as custom-base-url clients, which matches the compatibility pattern OpenRouter relies on. | 中 | SE001, SE035 |
| CE046 | AI SDK community docs describe OpenRouter as offering enterprise support, dedicated support, and custom SLAs for high-volume users. | 中 | SE032 |
| CE047 | Independent 2026 review coverage says OpenRouter is strong for early multi-model exploration but weaker on support speed and governed production controls. | 中 | SE036 |
| CE048 | The reviewed public materials in this chapter do not surface a standalone public SLA or response-time commitment document. | 低 | SE001, SE022, SE036 |
| CE049 | The reviewed public materials in this chapter do not include a public SOC 2, ISO 27001, or equivalent third-party audit artifact. | 低 | SE014, SE015, SE022 |
| CE050 | The reviewed public materials in this chapter do not document a public VPC or self-host deployment architecture for OpenRouter. | 低 | SE001, SE020, SE022, SE036 |
| CU001 | OpenRouter publicly frames itself as a single API and routing layer for both developers and enterprises. | 中 | SU001, SU002, SU011 |
| CU002 | Business Wire says the platform is used by more than 8 million global users that include AI-native startups and large enterprises. | 中 | SU002, SU003 |
| CU003 | OpenRouter’s Series B post says it serves more than 8 million developers across 400-plus models. | 高 | SU001, SU003 |
| CU004 | Workspaces documentation positions OpenRouter around projects, teams, and deployment stages inside one account. | 高 | SU005, SU006 |
| CU005 | The workspaces launch blog explicitly speaks to solo developers, enterprises shipping across multiple teams, and agents running staging and production environments. | 中 | SU006 |
| CU006 | Stripe Projects markets OpenRouter as a one-command provisioned service on the free plan, which supports a low-friction self-serve entry point for developers. | 中 | SU012 |
| CU007 | The API reference says OpenRouter normalizes request and response schemas across providers so customers only need to learn one interface. | 中 | SU011 |
| CU008 | OpenRouter says prompt retention is opt-in while request metadata such as tokens and latency is stored for reporting and ranking. | 中 | SU010 |
| CU009 | Service tiers let customers choose flex or priority processing, and the flex tier trades higher latency and lower availability for lower cost. | 中 | SU009 |
| CU010 | Workspaces overview says Enterprise accounts can set daily, weekly, monthly, or lifetime budgets per workspace. | 中 | SU005, SU007 |
| CU011 | The user-activity API exposes the last 30 completed UTC days of endpoint-level usage and supports organization-level user filters. | 中 | SU008 |
| CU012 | Workspaces overview says all organization members automatically access the Default workspace and admins manage access across workspaces. | 中 | SU005 |
| CU013 | Official and third-party Series B materials agree weekly OpenRouter traffic expanded from 5 trillion to 25 trillion tokens over the previous six months. | 高 | SU001, SU002, SU003 |
| CU014 | Business Wire and TechCrunch both cite about 100 trillion monthly tokens and more than 400 available models in mid-2026. | 高 | SU002, SU003 |
| CU015 | TechCrunch says enterprises and other AI users use OpenRouter to choose different models for different jobs to manage cost or improve task quality. | 中 | SU003 |
| CU016 | CapitalG describes OpenRouter as sitting between agents, applications, and the model ecosystem with routing permissions, spend visibility, and audit-friendly reporting. | 中 | SU004, SU002 |
| CU017 | OpenRouter now runs public LLM Rankings and App & Agent Rankings surfaces, showing ecosystem demand is strong enough to support public discovery products. | 中 | SU014, SU015 |
| CU018 | Sacra estimated OpenRouter grew from roughly $19 million annualized spend and $1 million annualized revenue at end-2024 to about $100 million annualized GMV and $5 million annualized revenue by May 2025. | 中 | SU026 |
| CU019 | The Series B post says OpenRouter’s recent roadmap was driven by the shift from experimentation into production apps and agents. | 中 | SU001 |
| CU020 | Workspaces, budgets, user-activity analytics, and service tiers show OpenRouter has added organization-facing controls beyond a raw model proxy. | 中 | SU005, SU007, SU008, SU009 |
| CU021 | Roo Code publishes a dedicated OpenRouter provider guide covering API keys, provider selection, and model setup. | 中 | SU016 |
| CU022 | Roo Code says it automatically fetches more than 100 OpenRouter models and adjusts cost calculations for BYOK’s 5% charge. | 中 | SU016 |
| CU023 | OpenHands has a dedicated OpenRouter provider page and its broader LLM guide recommends multiple `openrouter/...` model strings for agent workflows. | 中 | SU017, SU018 |
| CU024 | Aider documents OpenRouter-specific model strings and says many Aider users access Sonnet through OpenRouter. | 中 | SU019, SU020 |
| CU025 | Aider’s LLM overview says OpenRouter offers free access to many models with daily usage limits, which supports experimentation more than guaranteed enterprise throughput. | 中 | SU020, SU024 |
| CU026 | LiteLLM’s OpenRouter page supports production environment variables and treats OpenRouter as a first-class provider for text, chat, vision, and embedding workloads. | 中 | SU021 |
| CU027 | LiteLLM’s provider catalog places OpenRouter inside a larger OpenAI-compatible ecosystem, supporting the view that OpenRouter is often one layer inside another developer toolchain. | 中 | SU021, SU022 |
| CU028 | Product Hunt’s archived review page summarizes 86 reviews at 5.0 out of 5 and names involve.me, Clado, and Agents Base makers as praising OpenRouter. | 中 | SU023 |
| CU029 | The Product Hunt review evidence is strongest on maker advocacy and ease of integration, not on disclosed contract size or renewal depth. | 中 | SU023 |
| CU030 | None of the retained official or news sources disclose net revenue retention, gross retention, logo churn, average contract length, or renewal cohorts. | 低 | SU001, SU002, SU003, SU005 |
| CU031 | The clearest public durability proxies are workspace budgets, 30-day activity analytics, unified billing across workspaces, and explicit provider or service-tier controls. | 中 | SU005, SU006, SU007, SU008, SU009 |
| CU032 | TrueFoundry’s 2026 review says positive OpenRouter sentiment clusters among founders and builders in early product discovery, while production users raise support, security, and cost-control concerns. | 中 | SU024 |
| CU033 | TrueFoundry says its Trustpilot snapshot showed OpenRouter at 1.7 out of 5 across 41 reviews with 79% one-star ratings in May 2026. | 中 | SU024 |
| CU034 | TrueFoundry highlights agentic sessions that can consume credits quickly without a workflow-level stop, framing spend control as a production-scale risk. | 中 | SU024 |
| CU035 | TrueFoundry’s pricing analysis says enterprise buyers must negotiate SSO or SAML, SLAs, and support rather than rely on a universal public SLA package. | 中 | SU025 |
| CU036 | The same pricing analysis says OpenRouter is most attractive for prototyping or moderate multi-model use and becomes less compelling when teams need VPC-native governance or one-model high-volume economics. | 中 | SU025 |
| CU037 | OpenRouter’s February 2026 outage disclosure says customers saw both 500 errors and misleading 401 errors during two incidents lasting roughly 35 to 38 minutes. | 中 | SU013 |
| CU038 | OpenRouter says some customers spent time debugging their own API keys because infrastructure failures surfaced as 401s before the company switched that condition to 503s. | 中 | SU013 |
| CU039 | OpenRouter’s data-handling docs make procurement easier by offering opt-in prompt retention, but they also confirm metadata is still stored for reporting and ranking. | 中 | SU010 |
| CU040 | Workspaces overview says budgets are an Enterprise-plan feature, implying governance depth expands with plan tier rather than being uniform across the entire base. | 中 | SU005 |
| CU041 | Stripe Projects positions OpenRouter as a launch partner that can auto-provision accounts and keys, giving the company a distribution channel beyond direct website signup. | 中 | SU012 |
| CU042 | Sacra says power users bring OpenRouter keys into cloud IDEs such as Cline and Aider or tools like Fal.ai while paying inference on their own key. | 中 | SU026 |
| CU043 | Sacra says OpenRouter’s customer value proposition is one API, one dashboard for usage and billing, faster setup, and model switching without multiple provider SDKs. | 中 | SU026 |
| CU044 | The Roo Code, OpenHands, Aider, and LiteLLM docs show OpenRouter is deeply embedded in coding workflows, so partner-tool quality and upstream provider behavior directly shape user experience. | 中 | SU016, SU017, SU019, SU021, SU013 |
| CU045 | Public customer proof is materially thinner than overall usage signaling because most named evidence comes from developer-tool documentation, review summaries, or maker testimonials rather than enterprise case studies with contract detail. | 中 | SU016, SU017, SU019, SU021, SU023 |
| CU046 | Public sources do not disclose top-customer concentration, revenue mix by embedded partner versus direct accounts, or geography split by paying organization. | 低 | SU002, SU005, SU024, SU025 |
| CU047 | Business Wire and the Series B post both frame the market shift as movement from single-model pilots to multi-model production systems that need centralized control. | 中 | SU001, SU002 |
| CU048 | OpenRouter’s apps and rankings surfaces reinforce ecosystem pull, but those surfaces do not reveal how much revenue sits in any one customer, partner, or vertical. | 中 | SU014, SU015 |
| CR001 | OpenRouter terms say disputes with OpenRouter are resolved by binding arbitration and New York governing law. | 中 | SR001 |
| CR002 | OpenRouter terms say the company may suspend, restrict, disable, or terminate access to a model if a user may violate model terms or if a model provider requests action. | 中 | SR001 |
| CR003 | OpenRouter terms disclaim liability for model suspension, removal, degradation, or modification arising from model-provider terms or provider acts and omissions. | 中 | SR001 |
| CR004 | OpenRouter terms say users retain copyright in their inputs while output ownership depends on the model terms of the provider used. | 中 | SR001 |
| CR005 | OpenRouter privacy says the company does not control or take responsibility for how underlying LLMs handle inputs or outputs, including model training use. | 中 | SR002 |
| CR006 | OpenRouter privacy discloses use of Google Analytics to monitor and analyze site usage. | 中 | SR002 |
| CR007 | OpenRouter privacy says transfers outside the EEA or UK rely on European Commission adequacy decisions or standard contractual clauses. | 高 | SR002, SR026 |
| CR008 | OpenRouter's DPA frames OpenRouter as a processor or service provider handling personal data on behalf of customer controllers. | 中 | SR003 |
| CR009 | OpenRouter's DPA says customers can request information necessary to demonstrate compliance and contribute to audits. | 中 | SR003 |
| CR010 | OpenRouter's privacy guide says prompt retention on OpenRouter is always opt-in. | 中 | SR004 |
| CR011 | OpenRouter's provider-logging guide says customers can block routing to providers that may train on prompts. | 中 | SR005 |
| CR012 | OpenRouter's provider-logging guide says routing does not automatically change based on provider retention policies, so users must ignore providers that do not meet their own requirements. | 中 | SR005 |
| CR013 | OpenRouter's ZDR guide says some endpoints may retain data for abuse or legal reasons even if they do not train on that data. | 中 | SR006 |
| CR014 | OpenRouter's ZDR guide says the company itself does not retain prompts unless the customer specifically opts into prompt logging. | 高 | SR006, SR010 |
| CR015 | OpenRouter's ZDR guide says in-memory prompt caching is not treated as retention for ZDR purposes. | 中 | SR006 |
| CR016 | OpenRouter's Input & Output Logging feature stores prompts and completions in isolated Google Cloud Storage for at least three months and possibly longer unless customers request deletion. | 中 | SR010 |
| CR017 | OpenRouter Broadcast can send full request and response content, token counts, cost, provider, and tool-usage traces to external observability destinations. | 中 | SR020 |
| CR018 | OpenRouter's Sensitive Info Guardrail can redact or block matching inputs before requests reach a provider, but beta NLP detectors for names and addresses may allow a request through if the check times out. | 中 | SR009 |
| CR019 | OpenRouter's prompt-injection guardrail scans requests before forwarding and can flag, redact, or block a request with HTTP 403. | 中 | SR008 |
| CR020 | OpenRouter guardrails and workspace budgets let customers set spend limits that block requests once limits are exceeded. | 高 | SR007, SR019 |
| CR021 | OpenRouter's sovereign-AI guide says EU in-region routing is available only for enterprise customers and keeps decryption and processing within the EU. | 中 | SR021 |
| CR022 | OpenRouter's data-residency blog says customers should audit OpenRouter itself before routing regulated workloads because trust shifts from many providers to the routing layer. | 中 | SR022 |
| CR023 | European Commission AI Act guidance says GPAI obligations took effect in August 2025 and transparency obligations take effect in August 2026. | 高 | SR024, SR025 |
| CR024 | European Commission AI Act guidance says high-risk AI systems require risk management, logging, documentation, human oversight, robustness, and cybersecurity controls. | 中 | SR024 |
| CR025 | OpenRouter's human-oversight blog says deployers bear liability for consequential AI decisions and identifies August 2026 as the first hard deadline for relevant EU AI Act obligations. | 高 | SR023, SR024 |
| CR026 | The U.S. Copyright Office says a subsequent AI report will address model training, licensing, and liability, showing that core generative-AI IP questions remain unresolved. | 中 | SR027 |
| CR027 | OpenRouter's reliability guide says provider failover is automatic by default, while model-level fallbacks are opt-in. | 高 | SR014, SR011 |
| CR028 | OpenRouter's provider-selection guide says allow_fallbacks defaults to true and data_collection defaults to allow. | 中 | SR011 |
| CR029 | OpenRouter's model-fallbacks guide says fallback models can be triggered by rate limits, downtime, context-length errors, and moderation refusals. | 中 | SR012 |
| CR030 | OpenRouter's routing and reliability docs say narrowing the provider set with only, ignore, or order reduces fallback options and can trade reliability for control. | 高 | SR011, SR014 |
| CR031 | OpenRouter's outages blog says a third-party caching dependency caused the February 17 and 19, 2026 outages and that the company added circuit breakers, fallback caching, and more accurate 503 responses afterward. | 中 | SR013 |
| CR032 | OpenRouter said the February 17 outage began as roughly 20% failing requests and escalated into 80–90% failure rates once cache invalidation forced a thundering herd onto database lookups. | 中 | SR013 |
| CR033 | OpenRouter acknowledged that infrastructure lookup failures were misreported to customers as 401 authentication errors during the February outages. | 高 | SR013, SR017 |
| CR034 | OpenRouter Status recorded a separate 401-errors incident that affected both chat and generation APIs for about 1 hour and 21 minutes. | 中 | SR017 |
| CR035 | OpenRouter Status said Clerk authentication-provider degradation affected website logins and account access while API requests were unaffected. | 中 | SR015 |
| CR036 | OpenRouter Status recorded an Amazon Bedrock outage as a platform incident, showing that upstream-provider issues can surface at the OpenRouter layer. | 中 | SR016 |
| CR037 | OpenRouter Status said delayed API request logs also delayed budgets and billing events for nearly five hours. | 中 | SR018 |
| CR038 | OpenRouter's routing docs imply that strict data and provider restrictions can leave requests with fewer eligible paths when no compliant provider is available. | 中 | SR011, SR021, SR022 |
| CR039 | Amazon Bedrock now offers prompt routing across model families through a single endpoint to optimize quality and cost. | 中 | SR028 |
| CR040 | Microsoft Foundry's model router offers cost, balanced, and quality modes inside one deployment and auto-updates supported models over time. | 中 | SR029 |
| CR041 | OpenAI enforces rate limits and monthly usage limits at the organization and project level rather than only at the end-user level. | 中 | SR030 |
| CR042 | Google Cloud publishes region-by-region and model-by-model quotas for generative AI and agent-runtime request capacity. | 中 | SR031 |
| CR043 | OpenRouter's reliability guide says some 429 paths or partial-output failures have still consumed credits, so failed-request economics are not perfectly insulated. | 中 | SR014 |
| CR044 | OpenRouter's terms page links Stripe and Coinbase legal agreements into payment and credits-related functionality. | 中 | SR001, SR032, SR033 |
| CR045 | Stripe's services agreement uses binding arbitration and lets Stripe modify or discontinue aspects of its services subject to security, law, or provider obligations. | 中 | SR032 |
| CR046 | Coinbase's user agreement uses binding arbitration and says Coinbase may refuse, suspend, or terminate accounts or trading in its sole discretion. | 中 | SR033 |
| CR047 | OpenRouter's workspace budgets are enterprise-only, so strong spend-governance defaults are not universally available across the customer base. | 中 | SR019 |
| CR048 | Because OpenRouter sells reliability abstraction, a control-plane outage at OpenRouter is more thesis-damaging than a normal SaaS incident at a company whose product is not itself routing and failover. | 中 | SR013, SR014 |
| CR049 | OpenRouter's privacy and compliance controls are real, but many are opt-in, configuration-dependent, or enterprise-only, leaving residual exposure material for regulated traffic. | 中 | SR006, SR007, SR019, SR021, SR022, SR023 |
| CR050 | AWS and Microsoft shipping native routing, combined with upstream quota and policy control by model vendors, makes take-rate compression a more plausible long-term risk than raw traffic shortfall. | 中 | SR028, SR029, SR030, SR031 |
| CR051 | The public materials reviewed for this chapter do not disclose OpenRouter traffic or revenue concentration by model provider or enterprise customer. | 低 | SR011, SR021, SR022 |
| CR052 | The public materials reviewed for this chapter provide a DPA, routing controls, and AI-governance guidance, but they do not surface a fully diligence-ready certification and subprocessor pack. | 低 | SR003, SR021, SR022, SR023 |
| CR053 | The most important thesis-break triggers are repeated OpenRouter-originated outages, evidence that regulated traffic cannot stay within approved routing boundaries, or persistent billing-credit disputes. | 中 | SR013, SR018, SR021, SR022 |
| CR054 | Pre-investment diligence should request provider concentration, uptime/SLO history, security attestations, and billing-dispute metrics under NDA before underwriting regulated or mission-critical adoption. | 中 | SR013, SR018, SR021, SR022, SR023 |
| CV001 | OpenRouter announced a $113 million Series B in May 2026 led by CapitalG with participation from strategic and venture investors including NVentures, ServiceNow Ventures, MongoDB Ventures, and Menlo Ventures. | 高 | SV001, SV002 |
| CV002 | TechCrunch reported that OpenRouter’s May 2026 Series B priced the company at about $1.3 billion post-money. | 中 | SV003 |
| CV003 | TechCrunch said OpenRouter’s June 2025 Series A was $40 million at an estimated roughly $547 million post-money valuation. | 中 | SV003 |
| CV004 | OpenRouter’s official and BusinessWire materials say weekly volume reached 25 trillion tokens by May 2026, up 5x from 5 trillion tokens six months earlier. | 高 | SV001, SV002 |
| CV005 | OpenRouter’s official and BusinessWire materials say the platform serves over 8 million users across more than 400 models. | 高 | SV001, SV002 |
| CV006 | CapitalG describes OpenRouter as a single API layer that lets developers access, route, and optimize across hundreds of AI models while reducing lock-in and improving failover. | 中 | SV004 |
| CV007 | OpenRouter’s SEC Form D shows a 2025 exempt offering with a total offering amount of $47,606,863, total amount sold of $29,606,894, and first sale date of 2025-05-27. | 高 | SV005, SV006, SV007 |
| CV008 | The public SEC filing record reviewed for OpenRouter does not disclose current revenue, cash balance, burn, runway, or preferred-share economics. | 高 | SV005, SV006, SV007 |
| CV009 | If the reported $1.3 billion valuation is post-money and the full $113 million round is primary, the headline ownership sold is about 8.7% before any secondary component or option-pool changes. | 中 | SV001, SV003 |
| CV010 | Public sources reviewed do not reveal liquidation preferences, participation features, option-pool refresh, or secondary mix for the 2026 Series B. | 中 | SV001, SV003, SV005, SV006 |
| CV011 | Sacra estimated that OpenRouter hit $5 million of annualized revenue in May 2025 on $100 million of GMV, up from $1 million of annualized revenue at the end of 2024. | 中 | SV035 |
| CV012 | OpenRouter’s September 2025 fee update says non-crypto top-ups carry a 5.5% fee with a $0.80 minimum and crypto top-ups carry a 5.0% flat fee. | 中 | SV036 |
| CV013 | OpenRouter’s BYOK announcement says each customer gets 1,000,000 BYOK requests per month for free and pays a 5% fee beyond that threshold. | 中 | SV037 |
| CV014 | The public monetization mechanics reviewed point to a take-rate and platform-fee business layered on top of routed inference and BYOK traffic rather than a classic seat-based SaaS model. | 中 | SV035, SV036, SV037 |
| CV015 | ofox.ai argues that OpenRouter’s $0.80 minimum platform fee can turn small experimental top-ups into effective fees well above the stated 5.5% headline rate. | 中 | SV031, SV036 |
| CV016 | CostBench argues that OpenRouter’s true cost can exceed listed rates because routing does not always pick the cheapest provider and users can absorb additional markup or hidden usage friction. | 低 | SV032, SV033 |
| CV017 | The public file supports strong demand and usage growth but still does not disclose the current revenue run-rate required to confirm whether the $1.3 billion valuation is fundamentally supported. | 中 | SV001, SV002, SV003, SV035 |
| CV018 | As of July 2026, CompaniesMarketCap showed CoreWeave at roughly $47.52 billion of market value on about $5.13 billion of TTM revenue, or about 9.3x revenue. | 中 | SV008, SV009 |
| CV019 | As of July 2026, CompaniesMarketCap showed Cloudflare at roughly $88.42 billion of market value on about $2.16 billion of TTM revenue, or about 40.9x revenue. | 中 | SV010, SV011 |
| CV020 | As of July 2026, CompaniesMarketCap showed Datadog at roughly $93.84 billion of market value on about $3.67 billion of TTM revenue, or about 25.6x revenue. | 中 | SV012, SV013 |
| CV021 | As of July 2026, CompaniesMarketCap showed Snowflake at roughly $90.30 billion of market value on about $4.68 billion of TTM revenue, or about 19.3x revenue. | 中 | SV014, SV015 |
| CV022 | As of July 2026, CompaniesMarketCap showed GitLab at roughly $5.38 billion of market value on about $0.95 billion of TTM revenue, or about 5.7x revenue. | 中 | SV016, SV017 |
| CV023 | As of July 2026, CompaniesMarketCap showed Fastly at roughly $2.95 billion of market value on about $0.65 billion of TTM revenue, or about 4.5x revenue. | 中 | SV018, SV019 |
| CV024 | As of July 2026, CompaniesMarketCap showed Akamai at roughly $16.57 billion of market value on about $4.20 billion of TTM revenue, or about 3.9x revenue. | 中 | SV020, SV021 |
| CV025 | The public comparable set spans roughly 4x to 41x revenue, with the more relevant AI and control-plane names clustering between single digits and the mid-20s rather than consistently at 40x-plus. | 中 | SV008, SV009, SV010, SV011, SV012, SV013, SV014, SV015, SV016, SV017, SV018, SV019, SV020, SV021 |
| CV026 | Together AI officially announced an $800 million Series C in 2026. | 中 | SV025 |
| CV027 | Fireworks AI officially announced a $250 million Series C at a $4 billion valuation and said its annualized revenue had surpassed $280 million. | 中 | SV027 |
| CV028 | Fireworks AI’s prior official Series B announcement said the company raised $52 million at a $552 million valuation. | 中 | SV028 |
| CV029 | Fireworks AI’s Series C announcement says the company powers over 10,000 companies, serves hundreds of thousands of developers, and processes more than 10 trillion tokens per day. | 中 | SV027 |
| CV030 | Not Diamond markets itself as an intelligent model router for coding agents and claims 30%-plus cost savings with 2x faster developer cycles. | 低 | SV029 |
| CV031 | The private peer financing signals reviewed show capital is flowing to AI infrastructure companies that disclose larger scale or revenue markers than OpenRouter’s public file does today. | 中 | SV025, SV027, SV028, SV029 |
| CV032 | IDC argues that model routing matters not only for portability but also for optimizing performance, cost, and trust across fast-changing model options. | 中 | SV023 |
| CV033 | IDC says agentic AI systems increase the need for flexible routing across diverse models rather than one static model choice. | 中 | SV023 |
| CV034 | Research and Markets treats LLM gateway platforms as a distinct market with historic and forecast market-size analysis, TAM framing, and competitive-dynamics assessment through 2035. | 中 | SV022 |
| CV035 | TrueFoundry frames AI gateways as adding routing, fallbacks, guardrails, budgets, observability, and cost controls, while also implying that buyers can compare vendors on these practical dimensions. | 中 | SV034 |
| CV036 | At a $1.3 billion valuation, OpenRouter would need about $52 million of revenue to trade at 25x, about $72 million to trade at 18x, and about $144 million to trade at 9x revenue. | 中 | SV003, SV008, SV009, SV012, SV013, SV014, SV015 |
| CV037 | Using Sacra’s May 2025 $5 million annualized revenue estimate as the only public revenue anchor, the 2026 $1.3 billion valuation signal equates to about 260x that stale run-rate. | 中 | SV003, SV035 |
| CV038 | The bull case requires that current revenue has scaled far above the stale 2025 Sacra estimate and that the company can earn a premium multiple closer to high-growth AI infrastructure or software leaders. | 中 | SV001, SV002, SV019, SV020, SV021, SV027, SV035 |
| CV039 | The base case is that OpenRouter is a fast-growing but still brokerage-like control plane that deserves a mid-teens to mid-20s revenue multiple only after proving revenue quality, retention, and margin durability. | 中 | SV018, SV020, SV021, SV025, SV035 |
| CV040 | The bear case is that monetization remains closer to take-rate intermediation, hidden-cost friction limits conversion, and the market ultimately values the business closer to mid-single-digit infrastructure multiples. | 中 | SV022, SV023, SV031, SV032, SV033 |
| CV041 | Because current gross margin, net retention, customer concentration, and burn are undisclosed, the public evidence provides weak price support for a new investor at the current headline valuation. | 中 | SV001, SV003, SV005, SV006, SV035 |
| CV042 | A 3x gross outcome from a $1.3 billion entry implies roughly a $3.9 billion exit value before later dilution, which is easiest to justify only if OpenRouter reaches revenue scale much closer to premium private or public AI infrastructure peers. | 中 | SV003, SV018, SV020, SV021, SV027 |
| CV043 | The evidence set supports a research-more recommendation with medium confidence, a high risk rating, and a stretched valuation stance rather than a buy call at the headline price. | 中 | SV003, SV023, SV025, SV035 |
| CV044 | Even if the product is valuable, pricing and reliability complaints create a credible risk that some customers perceive markup without enough incremental performance or operational value. | 中 | SV031, SV032, SV038 |
| CV045 | Final underwriting should focus on current revenue, gross margin net of provider payouts, enterprise mix, net retention, top-customer concentration, and the exact preferred-term stack. | 中 | SV005, SV006, SV035, SV038 |
| CV046 | A thesis-break trigger is any evidence that revenue quality, retention, or margins look more like low-moat brokerage economics than durable infrastructure software economics. | 中 | SV035, SV031, SV032, SV033 |
| CV047 | A second thesis-break trigger is repeated reliability failure or competitive displacement by larger clouds and better-capitalized peers before OpenRouter proves enterprise stickiness. | 中 | SV023, SV027, SV029, SV038 |
| CV048 | If management can privately show that token growth has translated into high-margin recurring revenue above roughly $75 million with durable enterprise retention, the current valuation could move from stretched to arguable within a premium comp band. | 中 | SV001, SV002, SV020, SV021, SV027, SV035 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | OpenRouter | OpenRouter | The Unified Interface For LLMs. Better prices, better uptime, no subscriptions. |
| SO002 | OpenRouter | About - The Unified Interface For LLMs | OpenRouter | Started in early 2023 as the first LLM marketplace, OpenRouter has grown to become the largest and most popular AI gateway. |
| SO003 | OpenRouter | Enterprise AI Infrastructure Made Simple | OpenRouter | Unified access to 400+ AI models with zero operational overhead. One API, one bill, every AI provider. |
| SO004 | OpenRouter | Pricing | OpenRouter | Pricing plans for indie hackers, AI native startups, and enterprises: Free, Pay-as-you-go, Enterprise. |
| SO005 | OpenRouter | Privacy Policy | OpenRouter, Inc. respects your privacy and we are committed to protecting it through this Privacy Policy. |
| SO006 | OpenRouter | Terms of Service | OpenRouter’s address for Notice is: OpenRouter, Inc., 169 Madison Avenue, New York, NY 10016, United States. |
| SO007 | OpenRouter | OpenRouter Quickstart Guide | OpenRouter provides a unified API that gives you access to hundreds of AI models through a single endpoint, while automatically handling fallbacks and selecting the most cost-effective options. |
| SO008 | OpenRouter | Workspaces - Organize Projects, Teams, and Agents | Workspaces let you organize your OpenRouter projects into separate environments, each with its own API keys, routing defaults, guardrails, and observability. |
| SO009 | OpenRouter | Zero Data Retention - How OpenRouter gives you control over your data | OpenRouter has privacy settings that, when enabled, only allow you to route to endpoints that have a Zero Data Retention policy. |
| SO010 | OpenRouter | Data Collection - OpenRouter Privacy | OpenRouter does not store your prompts or responses, unless you opt in. |
| SO011 | OpenRouter | OpenRouter Raises $113M Series B — OpenRouter Blog | Today we’re announcing our $113M Series B, led by CapitalG. |
| SO012 | OpenRouter | Introducing Workspaces — OpenRouter Blog | We launched workspaces to organize your OpenRouter projects into separate environments, each with its own api keys, routing defaults, guardrails and observability. |
| SO013 | OpenRouter | OpenRouter Outages on February 17 and 19, 2026 — OpenRouter Blog | On February 17th and 19th, OpenRouter experienced related outages caused by failures in a third-party caching dependency. |
| SO014 | OpenRouter | OpenRouter Failover: Provider Failover vs Model Fallbacks Explained — OpenRouter Blog | Provider failover is automatic and on by default. Model fallbacks are opt-in. |
| SO015 | OpenRouter | Introducing the Unified Image API — OpenRouter Blog | Image generation on OpenRouter now has a dedicated API with unified access to 30+ models. |
| SO016 | OpenRouter | May Release Spotlight — OpenRouter Blog | We closed our $113M Series B, and we’re now routing 100 trillion tokens a month. |
| SO017 | OpenRouter Status | OpenRouter Status | Past Incidents. May 2026 to July 2026. No incidents reported. |
| SO018 | CapitalG | OpenRouter | Founded in 2023, the company sits between agents, applications and the model ecosystem, standardizing access to leading providers through one unified interface. |
| SO019 | Business Wire | OpenRouter Raises $113 Million CapitalG-led Series B as Weekly Volume Explodes to 25T Tokens | OpenRouter’s volume has surged to 25 trillion tokens per week (100 trillion tokens per month). |
| SO020 | TechCrunch | OpenRouter more than doubles valuation to $1.3B in a year | Popular AI gateway maker OpenRouter, founded in 2023, has raised a hefty $113 million Series B led by CapitalG. |
| SO021 | Morningstar | OpenRouter Raises $113 Million CapitalG-led Series B as Weekly Volume Explodes to 25T Tokens | OpenRouter, the AI model exchange, today announced a $113 million Series B led by Alphabet’s independent growth fund, CapitalG. |
| SO022 | The SaaS News | OpenRouter Raises $113M Series B | Founded in 2023 by Alex Atallah and Louis Vichy, OpenRouter is an AI model exchange platform. |
| SO023 | The AI Insider | OpenRouter Hits $1.3B Valuation After $113M Series B Led by Google’s CapitalG | OpenRouter has raised $113 million in a Series B round led by CapitalG ... lifting its valuation to approximately $1.3 billion. |
| SO024 | StreetInsider / SEC Filings mirror | Form D OpenRouter, Inc. | OpenRouter, Inc. ... DELAWARE ... 169 MADISON AVE #2404 ... NEW YORK ... Total Offering Amount $47,606,863 ... Total Amount Sold $29,606,894. |
| SO025 | TrueFoundry | OpenRouter Reviews 2026: Honest Verdict From Real Users | OpenRouter customer reviews flag support delays, account security concerns, free-tier rate limits, and agentic workloads that can quickly consume credits. |
| SO026 | Releasebot | OpenRouter Release Notes - June 2026 Latest Updates | OpenRouter launches the MCP server, giving coding agents live model rankings, pricing, docs, benchmarks, and test inference. |
| SO027 | WisdomAI | Inside Harvey AI’s $8B legal play and how OpenRouter ties LLMs together | Teams no longer want to be locked to a single LLM or provider. They want to route queries to the model that fits the task, cost, or compliance requirement. |
| SO028 | AI Market Watch | OpenRouter - AI Startup Profile | AI Market Watch | Target customer: Developers, startups, and companies building AI-powered applications who want to avoid vendor lock-in and easily switch between models. |
| SO029 | Tracxn (Wayback snapshot) | OpenRouter - 2025 Company Profile - Tracxn | OpenRouter is a funded company based in Anguilla, founded by Alex Atallah. OpenRouter has 1 institutional investor - Soma Capital. |
| SO030 | GitHub | OpenRouter | OpenRouterTeam/go-sdk ... Updated Jul 1, 2026. OpenRouterTeam/typescript-sdk ... OpenRouterTeam/python-sdk ... OpenRouterTeam/openrouter-examples. |
| SO031 | OpenRouter | Careers at OpenRouter | OpenRouter | We route billions of tokens every month ... Remote First ... Flexibility to work from anywhere in the US. |
| SM001 | OpenRouter | OpenRouter API Reference - Complete Documentation | OpenRouter normalizes the schema across models and providers so you only need to learn one. |
| SM002 | OpenRouter | Provider Routing - Smart Multi-Provider Request Management | By default, requests are load balanced across the top providers to maximize uptime. |
| SM003 | OpenRouter | Prompt Caching - Optimize AI Model Costs with Smart Caching | Sticky routing only activates when the provider’s cache read pricing is cheaper than regular prompt pricing. |
| SM004 | IDC | The future of AI is model routing | By 2028 70% of top AI-driven enterprises will use advanced multi-tool architectures to dynamically and autonomously manage model routing across diverse models. |
| SM005 | Research and Markets | AI Gateway Market - Global Forecast 2025-2032 - Research and Markets | The AI Gateway Market was valued at USD 3.21 billion in 2024 and is projected to reach USD 3.66 billion in 2025 ... reaching USD 9.61 billion by 2032. |
| SM006 | Intel Market Research | LLM Middleware Gateway Market Outlook 2026-2034 | The market is projected to grow from USD 18.9 million in 2026 to USD 189 million by 2034. |
| SM007 | The Business Research Company | Large Language Model (LLM) Gateway Platform Market Share, Size, Report 2035 | It will grow from $3.34 billion in 2025 to $4.23 billion in 2026 at a compound annual growth rate (CAGR) of 26.7%. |
| SM008 | The Business Research Company | Large Language Model Gateways Market Share, Size, Report 2026 | The large language model gateways market size has grown exponentially in recent years. It will grow from $2.18 billion in 2025 to $2.76 billion in 2026. |
| SM009 | Intel Market Research | AI API Gateway Market Outlook 2026-2034 | The market is projected to grow from USD 0.85 billion in 2026 to USD 2.12 billion by 2034. |
| SM010 | Research and Markets | Large Language Model (LLM) Router Market Report 2026 | Major trends include dynamic model routing, performance and cost optimization, load balancing and latency management, API management and integration, and monitoring and analytics dashboards. |
| SM011 | Cloudflare | Overview · Cloudflare AI Gateway docs | Cloudflare's AI Gateway allows you to gain visibility and control over your AI apps ... with features such as caching, rate limiting, request retries, model fallback, and more. |
| SM012 | Cloudflare | Pricing · Cloudflare AI Gateway docs | A 5% fee is applied to all credits purchased through Unified Billing ... Inference pricing from providers is passed through with no markup. |
| SM013 | Cloudflare | Caching · Cloudflare AI Gateway docs | This means caching is based on exact match of the entire request. |
| SM014 | Kong | Kong AI Gateway | Kong Docs | This normalized API layer provides multiple benefits ... request routing can be dynamic to optimize for cost, latency, or availability. |
| SM015 | Kong | AI Proxy - Plugin | Kong Docs | AI Proxy plugin accepts requests in standardized OpenAI formats, translates them to the configured target format, and then transforms the response back into a standard format. |
| SM016 | Kong | AI Prompt Guard - Plugin | Kong Docs | You can use a combination of allow and deny rules to preserve integrity and compliance when serving an LLM service using Kong Gateway. |
| SM017 | Amazon Web Services | Understanding intelligent prompt routing in Amazon Bedrock | Amazon Bedrock intelligent prompt routing provides a single serverless endpoint to efficiently route requests between different foundational models within the same model family. |
| SM018 | Amazon Web Services / Kong | Unlock Advanced AI Control with Kong AI Gateway And Amazon Bedrock | Amazon Web Services | A critical challenge that emerges at enterprise scale is the need for centralized AI governance and AI service consumer management. |
| SM019 | Microsoft | Model router for Microsoft Foundry concepts - Microsoft Foundry | Model router optimizes costs and latencies while maintaining comparable quality. |
| SM020 | Microsoft | AI gateway capabilities in Azure API Management | API Management also provides a unified model API (preview). It exposes multiple backends through a single OpenAI-compatible endpoint. |
| SM021 | LiteLLM | Router - Load Balancing | liteLLM | LiteLLM manages load-balance across multiple deployments ... and basic reliability logic - cooldowns, fallbacks, timeouts and retries. |
| SM022 | LiteLLM | Fallbacks | liteLLM | If a call fails after num_retries, fallback to another model group. |
| SM023 | TrueFoundry | A Definitive Guide to AI Gateways in 2026: Competitive Landscape Comparison | AI gateways unify access to many models behind one API, adding routing, fallbacks, guardrails, budgets, and observability. |
| SM024 | TrueFoundry | OpenRouter Vs AI Gateway: Differences, Use Cases & Best Choice | OpenRouter is ideal for experimentation and prototyping ... AI Gateways are necessary for enterprise-scale AI. |
| SM025 | Not Diamond | The Top 10 AI Gateways for the Multi-Model Future (2026) | OpenRouter is the de facto marketplace for LLMs ... 5.5% platform fee on inference means cost scales with volume. |
| SM026 | Maxim AI | Top 5 AI Gateways for Multi-Model Routing in 2026 | Done well, multi-model routing reduces token spend by 40-70% on mixed workloads while improving reliability through cross-provider failover. |
| SM027 | Maxim AI | Top 5 Enterprise LLM Gateways in 2026 | OpenRouter is a managed API service that provides access to hundreds of AI models from multiple providers through a single endpoint with unified billing. |
| SM028 | Google Cloud | Google models | Gemini Enterprise Agent Platform | Google Cloud Documentation | Our most cost-efficient model, optimized for low latency use cases for high-volume, cost-sensitive LLM traffic. |
| SP001 | Together AI | Together AI | The AI Native Cloud | Announcing our Series C. |
| SP002 | Together AI | Serverless models - Together AI docs | You call any supported model through a shared per-token API, with no provisioning, no replicas to size, and no minimum cost. |
| SP003 | Replicate | Replicate - Run AI with an API | Run and fine-tune models. Deploy custom models. All with one line of code. |
| SP004 | Replicate | Pricing – Replicate | You only pay for what you use on Replicate. Some models are billed by hardware and time, others by input and output. |
| SP005 | Replicate | Documentation – Replicate | |
| SP006 | Fireworks AI | Fireworks AI - Fastest Inference for Generative AI | Fireworks processes 30T+ tokens per day. |
| SP007 | Fireworks AI | Fireworks - Pricing | Serverless Inference ... per token pricing, zero setup and no cold starts. |
| SP008 | Fireworks AI | Build with Fireworks AI - Fireworks AI Docs | |
| SP009 | Martian | Martian: Understanding Intelligence | |
| SP010 | Martian | Introducing RouterBench | No single model can achieve optimal performance for all applications while remaining cost-effective. |
| SP011 | Martian | Martian Partners with Accenture, Launches Airlock Compliance for Enterprises | Martian builds model routers — systems that dynamically select the optimal AI model for each query. |
| SP012 | LiteLLM | LiteLLM | AI Gateway to provide model access, fallbacks and spend tracking across 100+ LLMs. All in the OpenAI format. |
| SP013 | LiteLLM | Getting Started | liteLLM | LiteLLM is an open-source library that gives you a single, unified interface to call 100+ LLMs. |
| SP014 | LiteLLM | Router - Load Balancing | liteLLM | |
| SP015 | GitHub | GitHub REST API: BerriAI/litellm | |
| SP016 | Amazon Web Services | Amazon Bedrock – Build genAI applications and agents at production scale – AWS | Amazon Bedrock powers generative AI for more than 100,000 organizations worldwide. |
| SP017 | Amazon Web Services | Understanding intelligent prompt routing in Amazon Bedrock - Amazon Bedrock | Amazon Bedrock intelligent prompt routing provides a single serverless endpoint to efficiently route requests between different foundational models within the same model family. |
| SP018 | Amazon Web Services | Amazon Bedrock Pricing – AWS | |
| SP019 | Microsoft | Microsoft Foundry documentation | Microsoft Learn | |
| SP020 | Microsoft | Model router for Microsoft Foundry concepts - Microsoft Foundry | Microsoft Learn | Model router ... delivers high performance while saving on costs, reducing latencies, and increasing responsiveness, while maintaining comparable quality, all packaged as a single model deployment. |
| SP021 | Microsoft | Microsoft Foundry - Pricing | Microsoft Azure | |
| SP022 | OpenRouter | Provider Routing - Smart Multi-Provider Request Management | By default, requests are load balanced across the top providers to maximize uptime. |
| SP023 | OpenRouter | Model Fallbacks - Automatic Failover Between Models | The models parameter lets you automatically try other models if the primary model’s providers are down, rate-limited, or refuse to reply due to content moderation. |
| SP024 | OpenRouter | Service Tiers - Control Cost and Latency Tradeoffs | Supported values are flex (lower cost, higher latency) and priority (faster, higher cost). |
| SP025 | OpenRouter | Pareto Router - Coding-Score-Based Model Selection | You express a single min_coding_score preference between 0 and 1, and the router routes your request to a coding model that meets that bar. |
| SP026 | OpenRouter | Fusion Router - openrouter/fusion | A panel of models answers your prompt in parallel, then a judge model compares their responses and returns structured analysis. |
| SP027 | OpenAI | Pricing | OpenAI API | |
| SP028 | Anthropic | Plans & Pricing | Claude by Anthropic | |
| SP029 | LlamaIndex | Routers | Developer Documentation | |
| SP030 | LangChain | LangChain overview - Docs by LangChain | Switch models with minimal code changes and keep your application portable as requirements evolve. |
| SI001 | OpenRouter | OpenRouter Raises $113M Series B — OpenRouter Blog | Over the last six months, weekly volume on OpenRouter has grown from 5 trillion to 25 trillion tokens. We are on pace to process over a quadrillion tokens this year and serve 8M+ developers building across 400+ models. |
| SI002 | Business Wire | OpenRouter Raises $113 Million CapitalG-led Series B as Weekly Volume Explodes to 25T Tokens | The company will use the new capital to expand its routing, governance, and optimization capabilities as enterprises increasingly deploy AI into production. |
| SI003 | TechCrunch | OpenRouter more than doubles valuation to $1.3B in a year | Popular AI gateway maker OpenRouter, founded in 2023, has raised a hefty $113 million Series B led by CapitalG... The New York Times reports that it landed at about $1.3 billion post-money. |
| SI004 | CapitalG | OpenRouter — CapitalG portfolio | Organizations can enforce controls like per-request data handling policies, team-level access and routing permissions, spend visibility, and audit-friendly usage reporting. |
| SI005 | OpenRouter | Get remaining credits | OpenRouter | Documentation | Get total credits purchased and used for the authenticated user. |
| SI006 | OpenRouter | Service Tiers - Control Cost and Latency Tradeoffs | Supported values are flex (lower cost, higher latency) and priority (faster, higher cost). |
| SI007 | OpenRouter | Stripe Projects - Add OpenRouter via Stripe CLI | OpenRouter ships with two plans through Stripe Projects: free (no credit card required) or pay-as-you-go (per-token usage pricing). |
| SI008 | OpenRouter | List workspace budgets | OpenRouter | Documentation | List of budgets configured for the workspace. |
| SI009 | OpenRouter | Query analytics data | OpenRouter | Documentation | Execute an analytics query with specified metrics, dimensions, filters, and time range. |
| SI010 | OpenRouter | Get user activity grouped by endpoint | OpenRouter | Documentation | Returns user activity data grouped by endpoint for the last 30 (completed) UTC days. |
| SI011 | OpenRouter | List all providers | OpenRouter | Documentation | The providers endpoint returns structured provider metadata such as status pages, datacenters, and terms links. |
| SI012 | OpenRouter | Latency and Performance | Minimizing Gateway Latency | To maintain accurate billing and prevent overages, OpenRouter performs additional database checks when a user’s credit balance is low. |
| SI013 | Securities and Exchange Commission | CIK0002073423 submissions JSON | "form":["D"], "filingDate":["2025-08-25"], "primaryDocument":["xslFormDX01/primary_doc.xml"] |
| SI014 | Securities and Exchange Commission | SEC FORM D | Total Offering Amount $47,606,863 USD ... Total Amount Sold $29,606,894 USD ... enter the total number of investors who already have invested in the offering: 30. |
| SI015 | StreetInsider | Form D OpenRouter, Inc. | Revenue Range ... X Decline to Disclose ... Type of Filing X New Notice Date of First Sale 2025-05-27. |
| SI016 | OpenRouter | Simplifying Our Platform Fee | Non-crypto payments: 5.5% of the order amount, with a minimum fee of $0.80. Crypto payments: 5.0% flat, no minimum fee. |
| SI017 | OpenRouter | 1 million free BYOK requests per month | Starting October 1st, every customer gets 1,000,000 “Bring Your Own Key” (BYOK) requests per month for free... requests will be charged at the usual rate of 5%. |
| SI018 | OpenRouter | Updates to our free tier — sustaining accessible AI for everyone | OpenRouter will directly cover some of the associated costs to ensure the most popular models remain freely accessible. |
| SI019 | OpenRouter | OpenRouter outages on February 17 and 19, 2026 | During the February 17th outage, approximately 20% of API requests failed... followed by 80-90% failure rates... The February 19th outage followed a similar pattern. |
| SI020 | OpenRouter | Introducing workspaces | We launched workspaces to organize your OpenRouter projects into separate environments, each with its own api keys, routing defaults, guardrails and observability. |
| SI021 | TrueFoundry | OpenRouter Pricing 2026: Plans, Costs, and Hidden Fees | The real cost becomes clearer at scale when credit purchase fees, BYOK charges, rate limits, missing public SLA terms, and governance gaps start to affect production decisions. |
| SI022 | TrueFoundry | OpenRouter Reviews 2026: Honest Verdict From Real Users | OpenRouter customer reviews flag support delays, account security concerns, free-tier rate limits, and agentic workloads that can quickly consume credits. |
| SI023 | Sacra | OpenRouter at $100M GMV | Sacra estimates that OpenRouter hit $5M in annualized revenue in May 2025, on $100M in GMV... OpenRouter customers buy tokens for inference... with tokens marked up roughly 5%. |
| SI024 | ofox.ai | OpenRouter Pricing 2026: Complete Model Cost Guide & Hidden Markup Breakdown | Budget for 5-7% overhead on top of whatever the model itself costs, and treat the calculator-quoted token rate as the floor, not the ceiling. |
| SI025 | CostBench | OpenRouter: Hidden Costs (2026) | The same model can be priced dramatically differently across providers on OpenRouter... Free models on OpenRouter are prone to timeouts, high-load failures, and require artificial pauses between requests. |
| SI026 | CostBench | OpenRouter Pricing 2026: 300+ LLM Models — Passthrough Rates | Free Models rateLimit: 20 req/min on free models... Pay-as-you-go billing: Prepay credits, minimum $5... 300+ models across all providers. |
| SE001 | OpenRouter | OpenRouter Quickstart Guide | |
| SE002 | OpenRouter | OpenRouter Models - Unified Access to 400+ AI Models | |
| SE003 | OpenRouter | Models | OpenRouter | |
| SE004 | OpenRouter | LLM Rankings | OpenRouter | |
| SE005 | OpenRouter | Auto Exacto - Automatic tool-calling provider optimization | Auto Exacto is a routing step that automatically optimizes provider ordering for all requests that include tools. |
| SE006 | OpenRouter | Auto Router - Intelligent Model Selection | |
| SE007 | OpenRouter | Response Caching - Cache Identical API Responses | |
| SE008 | OpenRouter | Tool & Function Calling - Use Tools with OpenRouter | |
| SE009 | OpenRouter | Server Tools - Model-Callable Tools by OpenRouter | |
| SE010 | OpenRouter | Web Search Server Tool - Real-Time Web Search for Any Model | |
| SE011 | OpenRouter | Web Fetch Server Tool - URL Content Retrieval for Any Model | |
| SE012 | OpenRouter | Structured Outputs - Type-Safe JSON Responses from AI Models | |
| SE013 | OpenRouter | Zero Completion Insurance - No Charge for Zero Token Responses | |
| SE014 | OpenRouter | Input & Output Logging - Privately Store Prompts and Completions | Prompt and response data is stored in an isolated Google Cloud Storage project with separate access controls. |
| SE015 | OpenRouter | Provider Logging - Provider Data Retention Policies | If you opt out of training in your account settings, OpenRouter will not route to providers that train. |
| SE016 | OpenRouter | Uptime Optimization - Ensure Reliable AI Model Access | |
| SE017 | OpenRouter | List all models and their properties | OpenRouter API Reference | |
| SE018 | OpenRouter | Preview the impact of ZDR on the available endpoints | OpenRouter API Reference | |
| SE019 | OpenRouter | Usage for Agents | OpenRouter Client SDKs | |
| SE020 | OpenRouter | Agent SDK Overview | OpenRouter | |
| SE021 | OpenRouter | Introducing the Unified Image API | Image generation on OpenRouter now has a dedicated API with unified access to 30+ models. |
| SE022 | OpenRouter | May Release Spotlight | Set per-member and per-key spend limits, lock traffic to a model and provider allowlist, enforce zero data retention, block prompt injection, and redact PII. |
| SE023 | OpenRouter | OpenRouter Failover: Provider Failover vs Model Fallbacks Explained | Provider failover is automatic and on by default. Model fallbacks are opt-in. |
| SE024 | OpenRouter | OpenRouter AI SDK Provider repo | |
| SE025 | OpenRouter | OpenRouter TypeScript SDK repo | |
| SE026 | OpenRouter | OpenRouter Agent SDK repo | |
| SE027 | OpenRouter | @openrouter/ai-sdk-provider package | There are 71 other projects in the npm registry using @openrouter/ai-sdk-provider. |
| SE028 | OpenRouter | @openrouter/sdk package | |
| SE029 | OpenRouter | openrouter Python SDK package | |
| SE030 | LangChain | ChatOpenRouter integration - LangChain | OpenRouter uses the OpenAI-compatible tool calling format. |
| SE031 | LiteLLM | OpenRouter provider - LiteLLM | |
| SE032 | AI SDK | OpenRouter Provider for the AI SDK | OpenRouter is a unified API gateway that provides access to hundreds of AI models from leading providers. |
| SE033 | OpenHands | OpenRouter - OpenHands Docs | |
| SE034 | Cline | OpenRouter - Cline | |
| SE035 | Pydantic | OpenAI-compatible providers - Pydantic AI Docs | |
| SE036 | TrueFoundry | OpenRouter Reviews 2026: Honest Verdict From Real Users | OpenRouter customer reviews flag support delays, account security concerns, free-tier rate limits, and agentic workloads that can quickly consume credits. |
| SU001 | OpenRouter | OpenRouter Raises $113M Series B — OpenRouter Blog | Over the last six months, weekly volume on OpenRouter has grown from 5 trillion to 25 trillion tokens. We are on pace to process over a quadrillion tokens this year and serve 8M+ developers building across 400+ models. |
| SU002 | Business Wire | OpenRouter Raises $113 Million CapitalG-led Series B as Weekly Volume Explodes to 25T Tokens | The platform is used by over 8 million global users, including AI-native startups and large enterprises. |
| SU003 | TechCrunch | OpenRouter more than doubles valuation to $1.3B in a year | The gateway helps enterprises and other AI users select different models for different jobs to control costs or increase reasoning and accuracy for the task at hand. |
| SU004 | CapitalG | OpenRouter | Organizations can enforce controls like per-request data handling policies, team-level access and routing permissions, spend visibility, and audit-friendly usage reporting. |
| SU005 | OpenRouter | Workspaces - Organize Projects, Teams, and Agents | Workspaces let you organize your OpenRouter projects into separate environments, each with its own API keys, routing defaults, guardrails, and observability. |
| SU006 | OpenRouter | Introducing Workspaces — OpenRouter Blog | Workspaces give you organization, flexibility, and control. |
| SU007 | OpenRouter | List workspace budgets | OpenRouter | Documentation | |
| SU008 | OpenRouter | Get user activity grouped by endpoint | OpenRouter | Documentation | |
| SU009 | OpenRouter | Service Tiers - Control Cost and Latency Tradeoffs | The example below requests the flex tier from OpenAI’s gpt-5 for a 50% discount in exchange for higher latency and lower availability. |
| SU010 | OpenRouter | Data Collection - OpenRouter Privacy | Any prompt retention on OpenRouter is always opt-in. |
| SU011 | OpenRouter | OpenRouter API Reference - Complete Documentation | |
| SU012 | OpenRouter | Stripe Projects - Add OpenRouter via Stripe CLI | OpenRouter is a launch partner, so you can add AI model access to any project with a single command. |
| SU013 | OpenRouter | OpenRouter Outages on February 17 and 19, 2026 — OpenRouter Blog | We know we let our customers down. |
| SU014 | OpenRouter | App & Agent Rankings | OpenRouter | |
| SU015 | OpenRouter | LLM Rankings | OpenRouter | |
| SU016 | Roo Code | Using OpenRouter With Roo Code | Roo Code Documentation | Roo Code automatically fetches all available models from OpenRouter's API (100+ models from various providers). |
| SU017 | OpenHands | OpenRouter - OpenHands Docs | OpenHands uses LiteLLM to make calls to chat models on OpenRouter. |
| SU018 | OpenHands | Overview - OpenHands Docs | These pages remain the authoritative provider references for both the Agent SDK and the OpenHands interfaces. |
| SU019 | Aider | OpenRouter | In particular, many aider users access Sonnet via OpenRouter. |
| SU020 | Aider | Connecting to LLMs | OpenRouter offers free access to many models, with limitations on daily usage. |
| SU021 | LiteLLM | OpenRouter | liteLLM | For production environments, you can dynamically configure the base_url using environment variables. |
| SU022 | LiteLLM | Providers | liteLLM | |
| SU023 | Product Hunt | OpenRouter Reviews | Product Hunt | The community submitted 86 reviews... Based on 86 reviews... 5.0. |
| SU024 | TrueFoundry | OpenRouter Reviews 2026: Honest Verdict From Real Users | OpenRouter reviews show a clear split between early-stage developers and production users. |
| SU025 | TrueFoundry | OpenRouter Pricing 2026: Plans, Costs, and Hidden Fees | Enterprise tier is custom-priced and adds SSO/SAML, contractual SLAs, priority support, and dedicated support. |
| SU026 | Sacra | OpenRouter at $100M GMV | Power users of AI models can 'bring their own key' to AI cloud IDEs like Cline & Aider. |
| SR001 | OpenRouter | Terms of Service | These Terms provide that all disputes between you and OpenRouter will be resolved by BINDING ARBITRATION. |
| SR002 | OpenRouter | Privacy Policy | We do not control, and are not responsible for, LLMs’ handling of your Inputs or Outputs, including for use in their model training. |
| SR003 | OpenRouter | Data Processing Agreement | OpenRouter will make available to Customer all information necessary to demonstrate compliance with the obligations of this DPA and allow for and contribute to audits. |
| SR004 | OpenRouter | Data Collection - OpenRouter Privacy | Any prompt retention on OpenRouter is always opt-in. |
| SR005 | OpenRouter | Provider Logging - Provider Data Retention Policies | OpenRouter does not have routing rules that change based on data retention policies of providers. |
| SR006 | OpenRouter | Zero Data Retention - How OpenRouter gives you control over your data | OpenRouter itself has a ZDR policy; your prompts are not retained unless you specifically opt in to prompt logging. |
| SR007 | OpenRouter | Guardrails - Organization Spending and Access Controls | Guardrails let organizations control how their members and API keys can use OpenRouter. |
| SR008 | OpenRouter | Prompt Injection Detection - Guardrail Regex Patterns | Block — The entire request is rejected with a 403 before it reaches the model. |
| SR009 | OpenRouter | Sensitive Info Guardrail - Automatic PII Detection and Redaction | The entire request is rejected with an HTTP 403 Forbidden error. |
| SR010 | OpenRouter | Input & Output Logging - Privately Store Prompts and Completions | Retention: Data is retained for a minimum of 3 months, and may be retained beyond 3 months at OpenRouter’s discretion unless you request deletion. |
| SR011 | OpenRouter | Provider Routing - Smart Multi-Provider Request Management | OpenRouter routes requests to the best available providers for your model. By default, requests are load balanced across the top providers to maximize uptime. |
| SR012 | OpenRouter | Model Fallbacks - Automatic Failover Between Models | The models parameter lets you automatically try other models if the primary model’s providers are down, rate-limited, or refuse to reply due to content moderation. |
| SR013 | OpenRouter | OpenRouter Outages on February 17 and 19, 2026 — OpenRouter Blog | A portion of users saw 500 or 401 errors on all API endpoints for 38 minutes starting at 5:27 AM UTC on February 17th, and for 35 minutes starting at 7:36 AM UTC on February 19th. |
| SR014 | OpenRouter | OpenRouter Failover: Provider Failover vs Model Fallbacks Explained — OpenRouter Blog | Some users have reported cases where error 429 consumed credits, or where partial outputs were counted despite an error. |
| SR015 | OpenRouter Status | Degraded website login | OpenRouter Status | Clerk, our authentication provider, is experiencing degraded performance. Logins and account access may be impacted while Clerk rolls out the identified fix. |
| SR016 | OpenRouter Status | Amazon Bedrock Outage | OpenRouter Status | Amazon Bedrock Outage. |
| SR017 | OpenRouter Status | 401 Errors across API surfaces | OpenRouter Status | Components impacted: Chat (/api/v1/chat/completions) and Generation (/api/v1/generation). |
| SR018 | OpenRouter Status | API Request Logs and Budget Enforcement are Delayed | OpenRouter Status | We are currently investigating an issue with delayed logging of API requests. This extends to delays in updating budgets and billing events. |
| SR019 | OpenRouter | Workspace Budgets - Per-Workspace Spend Limits | Workspace budgets are available on the Enterprise plan. |
| SR020 | OpenRouter | Broadcast - Send Traces to Observability Platforms | Each broadcast trace includes comprehensive information about your API request: Request & Response Data, Token Usage, Cost Information, Timing, Model Information, and Tool Usage. |
| SR021 | OpenRouter | Sovereign AI - In-Region AI Routing with OpenRouter | When enabled, your requests are guaranteed to only be decrypted within the designated region, and are only routed to providers operating in that region. |
| SR022 | OpenRouter | Enforce AI Data Residency at the Routing Layer — OpenRouter Blog | Before you route regulated workloads this way, audit OpenRouter’s own data handling policies and confirm the routing behavior matches your requirements. |
| SR023 | OpenRouter | EU AI Act & Colorado ADMT Compliance: Human Oversight for AI Agents — OpenRouter Blog | The first hard deadline lands in August 2026. |
| SR024 | European Commission | AI Act | The transparency rules of the AI Act will come into effect in August 2026. |
| SR025 | EU Artificial Intelligence Act | The Act Texts | EU Artificial Intelligence Act | The EU AI Act was published in the Official Journal (OJ) of the European Union on 12 July 2024. |
| SR026 | European Commission | Standard Contractual Clauses (SCC) | On 4 June 2021, the Commission issued modernised standard contractual clauses under the GDPR for data transfers. |
| SR027 | U.S. Copyright Office | Copyright and Artificial Intelligence, Part 2: Copyrightability Report | A subsequent part will turn to the training of AI models on copyrighted works, licensing considerations, and allocation of any liability. |
| SR028 | Amazon Web Services | Understanding intelligent prompt routing in Amazon Bedrock | Amazon Bedrock intelligent prompt routing provides a single serverless endpoint to efficiently route requests between different foundational models within the same model family. |
| SR029 | Microsoft | Model router for Microsoft Foundry concepts - Microsoft Foundry | Model router optimizes costs and latencies while maintaining comparable quality. |
| SR030 | OpenAI | Rate limits | OpenAI API | Rate limits are defined at the organization level and at the project level, not user level. |
| SR031 | Google Cloud | Generative AI on Gemini Enterprise Agent Platform quotas and system limits | This page provides a list of quotas by region and model, and shows you how to view and edit your quotas in the Google Cloud console. |
| SR032 | Stripe | Stripe Services Agreement - General Terms | Disputes between User and Stripe are subject to a class action waiver and will be resolved by individual binding arbitration. |
| SR033 | Coinbase | User Agreement - Coinbase | APPENDIX 5 INCLUDES AN AGREEMENT TO ARBITRATE WHICH REQUIRES, WITH LIMITED EXCEPTIONS, THAT ALL DISPUTES BETWEEN YOU AND US SHALL BE RESOLVED BY BINDING AND FINAL ARBITRATION. |
| SV001 | OpenRouter | OpenRouter Raises $113M Series B — OpenRouter Blog | |
| SV002 | BusinessWire | OpenRouter Raises $113 Million CapitalG-led Series B as Weekly Volume Explodes to 25T Tokens | |
| SV003 | TechCrunch | OpenRouter more than doubles valuation to $1.3B in a year | |
| SV004 | CapitalG | OpenRouter | |
| SV005 | Securities and Exchange Commission | OpenRouter, Inc. submissions JSON | |
| SV006 | Securities and Exchange Commission | SEC FORM D | |
| SV007 | StreetInsider | Form D OpenRouter, Inc. | |
| SV008 | CompaniesMarketCap | CoreWeave (CRWV) - Market capitalization | |
| SV009 | CompaniesMarketCap | CoreWeave (CRWV) - Revenue | |
| SV010 | CompaniesMarketCap | Cloudflare (NET) - Market capitalization | |
| SV011 | CompaniesMarketCap | Cloudflare (NET) - Revenue | |
| SV012 | CompaniesMarketCap | Datadog (DDOG) - Market capitalization | |
| SV013 | CompaniesMarketCap | Datadog (DDOG) - Revenue | |
| SV014 | CompaniesMarketCap | Snowflake (SNOW) - Market capitalization | |
| SV015 | CompaniesMarketCap | Snowflake (SNOW) - Revenue | |
| SV016 | CompaniesMarketCap | GitLab (GTLB) - Market capitalization | |
| SV017 | CompaniesMarketCap | GitLab (GTLB) - Revenue | |
| SV018 | CompaniesMarketCap | Fastly (FSLY) - Market capitalization | |
| SV019 | CompaniesMarketCap | Fastly (FSLY) - Revenue | |
| SV020 | CompaniesMarketCap | Akamai (AKAM) - Market capitalization | |
| SV021 | CompaniesMarketCap | Akamai (AKAM) - Revenue | |
| SV022 | Research and Markets | Large Language Model (LLM) Gateway Platform Market Report 2026 | |
| SV023 | IDC | The future of AI is model routing | |
| SV024 | Together AI | Together AI | The AI Native Cloud | |
| SV025 | Together AI | Announcing our $800M Series C to accelerate the shift to open-source AI | |
| SV026 | Fireworks AI | Fireworks AI - Fastest Inference for Generative AI | |
| SV027 | Fireworks AI | Fireworks AI Raises $250M Series C to Power the Future of Enterprise AI | |
| SV028 | Fireworks AI | Fireworks AI Raises $52M Series B to Lead Industry Shift to Compound AI Systems | |
| SV029 | Not Diamond | Not Diamond - Model Routing for Coding Agents | |
| SV030 | Not Diamond | About | |
| SV031 | ofox.ai | OpenRouter Pricing 2026: Complete Model Cost Guide & Hidden Markup Breakdown | |
| SV032 | CostBench | OpenRouter: Hidden Costs (2026) | |
| SV033 | CostBench | OpenRouter Pricing 2026: 300+ LLM Models — Passthrough Rates | |
| SV034 | TrueFoundry | A Definitive Guide to AI Gateways in 2026: Competitive Landscape Comparison | |
| SV035 | Sacra | OpenRouter at $100M GMV | |
| SV036 | OpenRouter | Simplifying Our Platform Fee — OpenRouter Blog | |
| SV037 | OpenRouter | 1 million free BYOK requests per month — OpenRouter Blog | |
| SV038 | OpenRouter | OpenRouter Outages on February 17 and 19, 2026 — OpenRouter Blog |