LMArena

1.1 Identity and Product

LMArena, now operating under the brand "Arena," is an AI model evaluation platform that enables users to compare frontier AI models through anonymous head-to-head battles. A user submits a prompt to two unlabeled models simultaneously, votes on the better response, and only then sees which models were judged. Aggregated votes power a public leaderboard using a Bradley-Terry / Elo-based ranking algorithm, generating continuous, real-world performance comparisons across large language models and multimodal AI systems. The corporate entity, Arena Intelligence Inc., was incorporated on April 18, 2025, transitioning the Chatbot Arena research project into a commercial enterprise. Headquarters is in the San Francisco Bay Area (specific address undisclosed), consistent with UC Berkeley origins. Primary websites operate at arena.ai and lmarena.ai. LMArena's commercial product, AI Evaluations, launched in September 2025, providing enterprises, model labs, and developers with community-grounded performance analytics, representative feedback samples, and service-level agreements. Named commercial customers include OpenAI, Google, and xAI. The business model is evaluation-as-a-service: clients pay for systematic assessments of model performance across domains such as software engineering, law, and medicine. The platform has expanded well beyond text-only LLM comparison to include Search Arena, WebDev Arena, Vision Arena, text-to-image, text-to-video, and Agent Arena (launched June 2026). This multimodal expansion broadens LMArena's evaluable surface and addressable customer base. As of June 2026, the Arena brand is the de facto public leaderboard for frontier AI models, with its rankings cited in product announcements, investor materials, and academic publications by every major AI lab. [CO001, CO002, CO005, CO007, CO008, CO009]

1.2 Founders and Leadership

LMArena was co-founded by three people with UC Berkeley affiliations. Anastasios Angelopoulos, CEO, was a postdoctoral researcher in Statistics and Machine Learning at UC Berkeley and is the primary public spokesperson. Wei-Lin Chiang, co-founder, completed his PhD in distributed systems at Berkeley and was instrumental in building the FastChat serving framework that underlies Chatbot Arena. Ion Stoica, a UC Berkeley Computer Science professor and serial entrepreneur (co-founder of Databricks and Anyscale), joined as co-founder, providing commercialization experience and industry credibility. Original Chatbot Arena research involved additional contributors including Lianmin Zheng and faculty advisors Michael Jordan and Joseph Gonzalez. The founding team demonstrates strong founder-market fit: Angelopoulos and Chiang co-authored the key academic papers establishing the evaluation methodology, and Stoica's track record at Databricks provides operational depth rarely seen in academic spinouts. Key-person dependency is elevated: Angelopoulos serves as both scientific lead and CEO. No non-founder C-suite executives have been publicly named. No board composition or independent director disclosures have been made, typical for early-stage private companies. No leadership departures or governance changes have been disclosed since incorporation. However, the 2025 benchmark integrity controversy revealed a tension between Stoica (who publicly disputed the Leaderboard Illusion findings) and external researchers, highlighting how the founding team's credibility is central to the platform's legitimacy. [CO003, CO004, CO005, CO006, CO021, CO030]

1.3 Funding History and Investors

LMArena's pre-commercial funding came from grants and donations, including contributions from Google's Kaggle platform, Andreessen Horowitz, and Together AI, primarily in the form of compute resources and cash to support research infrastructure. In May 2025, following incorporation, LMArena raised a $100 million seed round co-led by Andreessen Horowitz and UC Investments (the University of California's endowment), at a $600 million post-money valuation. Lightspeed Venture Partners, Felicis, and Kleiner Perkins also participated. In January 2026, the company raised a $150 million Series A at a $1.7 billion post-money valuation — nearly triple the seed valuation — co-led by Felicis and UC Investments, with Andreessen Horowitz, The House Fund, LDVP, Kleiner Perkins, Lightspeed, and Laude Ventures participating. Total capital raised stands at approximately $250 million as of January 2026. UC Investments' dual role as University of California endowment manager and institutional backer of a UC Berkeley spinout, and Andreessen Horowitz's portfolio overlap with models on the Arena leaderboard (including Mistral), represent potential conflicts of interest noted by external critics. No secondary transactions or debt facilities have been publicly disclosed. [CO010, CO011, CO012, CO013, CO014, CO021]

1.4 Traction Metrics and Financial Indicators

By January 2026, LMArena reported over 5 million monthly users across 150 countries and 60 million conversations per month. The community accumulated more than 50 million votes across text, vision, web development, search, video, and image modalities, and participated in evaluating more than 400 distinct AI models. LMArena also released 145,000 open-source battle data points from expert and occupational evaluation categories. On the financial side, the annualized consumption run rate — the company's ARR proxy — surpassed $30 million in December 2025, approximately four months after commercial launch. This is stated as an annualized run rate rather than realized GAAP revenue, and independently audited figures are unavailable. The rapid ramp from zero to $30 million annualized in four months is exceptional, but the structure of the commercial service (SLA-bound deliverables with community raters) implies potential labor cost constraints on gross margin that have not been disclosed. LMArena has not disclosed headcount, burn rate, or gross margin, which limits financial diligence from publicly available information alone. Deduplication filters remove approximately 10% of submitted votes, and identity-leak detection removes fewer than 4% of all votes as of July 2025 methodology updates, indicating active data quality management. [CO015, CO017, CO018, CO019, CO020, CO023]

1.5 Milestones and Adverse Events

LMArena's trajectory spans approximately three years from research demo to unicorn. The original Chatbot Arena launched in May 2023. Academic publications followed rapidly, culminating in the Chatbot Arena paper (arXiv:2403.04132) establishing its Bradley-Terry methodology. By 2024 the platform had become the de facto reference benchmark, cited by every major AI lab. The first material adverse event came in early 2025 when Meta tested at least 27 private Llama 4 model variants on Chatbot Arena, submitting an arena-optimized version that scored near the top while the publicly released version ranked 32nd. LMArena apologized and updated leaderboard policies. In April 2025, a peer-reviewed paper titled "The Leaderboard Illusion" (Cohere, Stanford, MIT, Ai2) formally documented alleged systematic bias in LMArena's private testing practices. LMArena co-founder Ion Stoica publicly disputed the findings as "inaccuracies." LMArena introduced new sampling algorithms and published updated transparency policies in response. The company incorporated in April 2025, raised $100 million in May 2025, launched its commercial product in September 2025, and closed the Series A in January 2026, achieving unicorn status. No regulatory investigations, litigation, sanctions, or enforcement actions have been publicly disclosed as of June 2026. [CO001, CO002, CO005, CO016, CO027, CO028]

1.6 Exhibits

2.1 Market Boundary and Definition

LMArena should be analyzed as part of the AI evaluation and benchmarking software layer rather than as a general model-infrastructure or observability company. The included market spans human-preference benchmarking, automated regression testing, domain-specific evaluation workflows for regulated or expertise-heavy use cases, and newer agent-evaluation products that score multi-step task success instead of only single-turn chat quality. Excluded from the serviceable market are foundation-model training compute, generic MLOps orchestration, broad developer tooling, and open-source evaluation scripts that are not sold as managed workflow software. That boundary matters because some publishers size a broad evaluation-platform category, while others size a narrow benchmarking-tools niche, producing multi-fold TAM divergence. LMArena’s own commercial messaging emphasizes law, medicine, and engineering evaluation, which suggests its monetizable market is defined more by high-stakes validation spend than by public leaderboard traffic alone.[CM006, CM007, CM008, CM009, CM016, CM018]

2.2 Market Sizing and Contested Estimates

The cleanest broad 2026 market anchor in the verified source set is the AI model evaluation platform market report showing $1.86B in 2025 growing to $2.36B in 2026 at 27.3% CAGR, and $6.24B by 2030. That lens likely includes enterprise evaluation software sold across model labs, application developers, and compliance-heavy organizations. A narrower lens from Precedence Research implies a materially smaller benchmarking-tools segment, around $0.85B in 2026, because it appears to exclude some bundled or adjacent workflows and emphasizes evaluation plus benchmarking tools rather than the full platform layer. Gartner’s 2026 AI spending forecast and Presenc AI’s production-adoption survey both support the idea that evaluation budgets can expand quickly from here, but they do not resolve the boundary problem. The practical conclusion is that LMArena’s real monetizable market is likely much smaller than the broad TAM yet still large enough to support a meaningful standalone company if it captures trusted, workflow-embedded spend.[CM001, CM002, CM003, CM004, CM005, CM021]

2.3 Buyer Segmentation and Procurement

LMArena’s buyer map has two distinct centers of gravity. First are frontier labs and model API vendors that need credible third-party benchmarking, launch validation, and competitive signaling before or after major releases. Second are regulated or expertise-heavy enterprises that need domain-specific evaluation for legal, medical, and engineering workflows where failure costs are high and public consumer benchmarks are insufficient. These segments likely purchase differently: labs fund evaluation from central model, safety, or research-platform budgets, while enterprises often buy through AI-platform leaders, product owners, or risk and quality teams. Competition is also mixed. Scale AI, Arize, Galileo, Patronus, MLflow, and large cloud or platform vendors all attack adjacent pieces of the stack, and CoreWeave’s acquisition of Weights & Biases shows that experiment, observability, and evaluation workflows are increasingly converging in enterprise procurement.[CM010, CM011, CM012, CM013, CM014, CM015]

2.4 Growth Drivers and Adoption Constraints

Several forces support above-market growth for trusted evaluation vendors. Global AI software and model spending are surging in 2026, enterprise AI has moved into production at a high share of large companies, and the shift toward agents increases the need for workflow-level evaluation rather than static prompt testing. Benchmark rivalry itself also stimulates demand, because model providers want external proof points and enterprises want independent quality signals. Yet the category also has real constraints. Benchmark gaming allegations, questions about crowdsourced-rater bias, and skepticism about public leaderboards all undermine willingness to pay for scores that are not clearly tied to production outcomes. Standalone vendors may also face pricing pressure from bundled cloud and MLOps tools, while limited public pricing disclosure makes it hard to separate durable software demand from novelty-driven experimentation.[CM026, CM027, CM028, CM029, CM030, CM031]

2.5 Evidence Gaps and Contradictions

The main diligence problem is not source scarcity but source mismatch. Public market reports disagree on the category boundary; company and media sources disclose valuation and anecdotal customer names but not enough contract detail to model share; competitor-revenue data are fragmentary; and no verified source in this run discloses what percentage of frontier-lab or enterprise AI budgets is actually spent on evaluation software. That means the chapter can support a credible range-based view but not a single precise SAM or market-share conclusion. Investors should preserve the contradiction instead of forcing a point estimate: LMArena may already be large relative to a narrow benchmarking-tools niche, while still tiny relative to the broader AI evaluation platform opportunity. Resolving that gap requires customer, pricing, and budget-intensity evidence that is not public in the verified source set.[CM003, CM024, CM038, CM039, CM040, CM041]

2.6 Exhibits

3.1 Competitive landscape: direct benchmarkers, automated leaderboards, data labelers, and build-it-yourself substitutes

LMArena operates in the AI model evaluation and benchmarking space, where no single competitor replicates its full stack. The addressable competitive set has four tiers. First, human-preference platforms: LMArena is the dominant public platform for crowdsourced pairwise model comparison; no comparable free alternative exists at anything close to its 5 million monthly user scale or 60 million monthly conversation volume as of January 2026. Second, automated academic leaderboards: the HuggingFace Open LLM Leaderboard (powered by EleutherAI lm-evaluation-harness), Stanford HELM, and BenchLM aggregate standardized benchmark scores such as MMLU, GPQA Diamond, and SWE-Bench across hundreds of models without human preference voting. These tools are free and open-source but measure task accuracy, not holistic user satisfaction. Third, enterprise evaluation vendors: Scale AI's GenAI Platform offers customized evaluation, fine-tuning, and data-labeling pipelines for enterprise and government clients at $93 K–$400 K+ per engagement; it does not host a public leaderboard. Fourth, status-quo substitutes: AI labs can build internal evaluation pipelines using EleutherAI's harness or OpenAI Evals and run their own test sets, avoiding third-party dependency entirely. Likely entrants include any major cloud provider building a benchmark surface to differentiate its model marketplace, or a large AI lab spinning up an in-house neutral evaluation arm. The meaningful competitive dimensions are human-preference scale, enterprise revenue potential, third-party independence, and methodology credibility.[CP001, CP002, CP003, CP011, CP012, CP014]

3.2 Feature and capability comparison: human preference versus automated benchmarks

The core capability divide in AI evaluation is between human-preference leaderboards (LMArena, its predecessors at LMSYS) and automated benchmark pipelines (HuggingFace Open LLM Leaderboard, Stanford HELM, EleutherAI lm-evaluation-harness, OpenAI Evals). LMArena's Elo-based ranking system, described in the 2024 Chatbot Arena paper, aggregates user-submitted pairwise votes across any task the user brings, creating a continuously updated signal that reflects real usage patterns rather than curated test sets. In practice this means LMArena is more sensitive to style, fluency, and helpfulness signals that users perceive in everyday tasks — coding, writing, answering questions — but less reproducible than automated benchmarks because two users can legitimately prefer opposite outputs. Automated leaderboards are reproducible and task-specific. Stanford HELM evaluates across accuracy, calibration, robustness, bias, and efficiency dimensions on a common harness, favoring researchers who need controlled measurements. The HuggingFace Open LLM Leaderboard tracks primarily open-source and open-weight models on standardized academic tests. EleutherAI's harness underlies both and is freely forkable. BenchLM aggregates 249 benchmarks across 261 models as of June 2026 and separately tracks price and speed. ArtificialAnalysis provides independent API throughput, latency, and cost benchmarking. Scale AI's enterprise evaluation product fills a different gap: custom, private, SLA-backed evaluation against a client's specific production use case rather than a public leaderboard — a model that does not compete with LMArena for brand recognition but may compete for enterprise budget. LMArena's key differentiators are network effects (more users → more votes → more statistically stable rankings), cross-modal coverage (text, vision, image generation, video, web dev as of 2025–2026), and the breadth of pre-release partnerships with frontier labs. Its weaknesses are methodological reproducibility, user-base bias toward technically sophisticated early adopters, and the inherent tension between selling evaluation services to the same labs it ranks. No competitor combines the public leaderboard surface with the enterprise paid evaluation service in a single brand, which is both LMArena's structural moat and its conflict-of-interest risk.[CP006, CP007, CP008, CP009, CP013, CP015]

3.3 Moat durability, switching costs, and adverse competitive evidence

LMArena's competitive durability rests on three factors: a crowdsourced preference dataset that is the largest of its kind, community trust reinforced by academic roots (UC Berkeley LMSYS Org, FastChat open-source infrastructure), and deep citation network effects — frontier labs cite Arena rankings in marketing materials and investor communications, raising the cost of defection. Switching costs for AI labs are real: if Arena ceases to be the market's signal, their existing benchmark investments lose marketing value. This creates a mutual dependency between LMArena and its client labs that provides revenue stability but also structural independence risk. The adverse evidence on moat is credible and material. The 2025 Leaderboard Illusion paper by Singh et al. (Cohere, Stanford, MIT, Ai2) quantifies data access asymmetry: Google and OpenAI received an estimated 19.2 % and 20.4 % of total Arena data respectively, while 83 combined open-weight models received only 29.7 %. Meta tested 27 private model variants before the Llama 4 release, selecting only the top scorer, which TechCrunch confirmed then ranked 32nd once the vanilla Maverick was re-submitted. LMArena denied the characterization of bias but committed to algorithm changes, signaling the critique had operational merit. A second adverse dynamic is Goodhart's Law: as Arena rankings drive procurement decisions, labs will optimize for Arena-specific patterns, degrading the signal quality and increasing the likelihood a well-resourced competitor can credibly claim superior methodology. The Elo algorithm underlying Arena is open-source and replicable by any well-funded team. What is not easily replicated is the community. Scale AI has a larger enterprise footprint and deeper data-labeling expertise, but has not built a public preference leaderboard. The real displacement threat is internal build: an AI lab that decides Arena's conflict-of-interest problem is unsolvable could fund a competing neutral body. The free academic tools (HELM, HuggingFace) already serve that credibility function for research use cases, keeping a ceiling on how much LMArena can claim scientific neutrality.[CP028, CP029, CP030, CP034, CP035, CP036]

3.4 Exhibits

4.1 Revenue model, pricing, and commercial traction

LMArena's revenue model is a freemium-to-enterprise funnel. The free public leaderboard — which serves 5 million monthly users across 150 countries and generates 60 million model-comparison conversations per month — acts as the community and trust-building layer. Enterprises, model labs, and AI developers pay for LMArena's commercial AI Evaluations product (launched September 2025), which provides custom evaluation panels grounded in real user feedback, representative data samples, and SLA-committed delivery timelines. LMArena has not published a public price list; all enterprise contracts are negotiated directly. Three confirmed enterprise client categories are AI labs (OpenAI, Google, xAI cited explicitly in the January 2026 Series A press release), software enterprises, and regulated professional verticals (law, medicine, scientific research). The company's annualized consumption run rate surpassed $30 million in December 2025 — a pace set less than four months after product launch. Revenue recognition is an important caveat: LMArena and TechCrunch both describe this figure as a "consumption run rate," not a realized annual figure. TechCrunch notes the company "describes its annual recurring revenue (ARR)" as a consumption rate, reflecting usage-based billing rather than upfront contracted SaaS ARR. With approximately 100 enterprise clients (per GetLatka aggregation), the implied average contract value is approximately $300,000 per client — an estimate derived from dividing the run rate by the client count rather than a company-disclosed figure. Actual ACV distribution, contract duration, and renewal rates are not publicly available.[CI001, CI002, CI003, CI004, CI005, CI006]

4.2 Cost structure, unit economics, and capital adequacy

LMArena's cost structure has three primary categories inferred from public evidence. First, compute and infrastructure: serving 60 million conversations per month across frontier models requires substantial cloud compute. While the public leaderboard uses model APIs from partner labs (which may be provided at reduced or no cost under partnership agreements), the enterprise evaluation product likely incurs real inference costs at scale. Second, engineering and research headcount: the company had 41 employees as of the January 2026 Series A announcement. At typical Silicon Valley all-in compensation for a technical startup, this implies annual headcount costs of roughly $15–25 million. Third, community management, data annotation quality assurance, and go-to-market costs that are impossible to size from public information. Gross margins are not disclosed; a software-like model with low marginal cost per evaluation could achieve 60–80% gross margins, but high compute costs per conversation could compress this to 40–60%. No third-party estimate with strong foundation exists. Capital adequacy appears strong in the near term. The company has raised $250 million total across its May 2025 seed round ($100 million at $600 million valuation) and January 2026 Series A ($150 million at $1.7 billion valuation). With a lean team of 41 employees and first commercial revenue only established in September 2025, the cash burn is likely well below $30 million annually as of early 2026. Rough estimates suggest a runway of at least 18–36 months without further fundraising, but no official burn rate or cash position has been disclosed. LMArena stated it will use the Series A funds to expand its technical team, strengthen research capabilities, and build new platform features — all of which will increase future burn. The Recall Capital-LMArena feeder fund (SEC Form D, filed February 2026) confirms the capital-formation activity but does not disclose LMArena's own balance sheet. No debt or project-finance obligations have been reported publicly.[CI009, CI010, CI011, CI012, CI013, CI014]

4.3 Financial risks, conflict of interest, and diligence blockers

The most material financial risk is structural: LMArena earns revenue from the same AI labs — OpenAI, Google, xAI — that it publicly ranks on its leaderboard. CTOL Digital documented this paradox explicitly after the January 2026 funding: the $1.7 billion valuation is "at 57 times that run rate, pricing in not just growth, but the assumption that this inherent conflict can be managed indefinitely." If enterprise clients come to believe that LMArena's rankings are influenced by commercial relationships (as the Leaderboard Illusion paper alleged for data-access asymmetry), the evaluation service's pricing power erodes and the free leaderboard's trust advantage — which is the core marketing asset — collapses simultaneously. This dual exposure is unusual: most SaaS companies face customer churn risk; LMArena faces credibility churn risk where the same event that loses a client also damages the free product that generates new clients. A second risk is revenue concentration: with approximately 100 enterprise clients and a run-rate dominated by three major AI labs, any single-client defection from OpenAI, Google, or xAI could materially reduce revenue. A third risk is the "consumption rate" framing: if billing is usage-based, revenue in any given month may not be predictable, and the $30 million annualized figure reflects a peak-month extrapolation rather than a contracted forward obligation. Fourth, the valuation at 57x run rate implies the capital markets are pricing in both high revenue growth (revenue must exceed $100 million to justify a 17x multiple, a common late-stage benchmark) and durable gross margins — both of which are unverifiable from public data. Any slowdown in enterprise adoption or loss of methodological trust could trigger a significant valuation reset that complicates future fundraising or exit optionality.[CI021, CI022, CI023, CI024, CI025, CI032]

4.4 Exhibits

5.1 Product Portfolio and Arena Modalities

LMArena's core user-facing product is the Arena platform at arena.ai, which presents battles between two anonymized AI models for a given prompt and records the community's preference vote. As of June 2026 the platform supports ten distinct evaluation arenas: Text (general chat), Code (agentic web-development and coding), Search (retrieval-augmented generation), Agent (autonomous multi-step task completion), Vision (multimodal understanding), Text-to-Image (image generation), Text-to-Video (video generation), Image Edit (image-editing models), Document (long-form document analysis), and a separate Direct Chat mode where users interact with a chosen model without the pairwise battle structure. The Agent Arena is the newest and most technically complex arena, launched June 4, 2026, which ranks orchestrator models on causal treatment effects derived from 2M+ tool calls per week. Code Arena, rebuilt from the earlier WebDev Arena (launched December 2024), operates as a live, isolated coding environment where models autonomously create, modify, and execute files through structured tool calls, with outputs persisted in Cloudflare R2 and displayed via CodeMirror 6. Since May 2026 LMArena also includes "Battles in Direct," routing 10% of Direct Chat sessions into anonymous pairwise battles to increase daily vote volume. The commercial product, AI Evaluations (launched September 2025), is a contract evaluation service sold to AI labs and enterprises, offering in-depth evaluations with representative feedback samples and SLA-backed delivery timelines. All public model evaluations are available free of charge on the public leaderboard; the commercial product adds private, confidential, SLA-governed evaluation runs. [CE001, CE002, CE008, CE012, CE029, CE032]

5.2 Technical Architecture and Evaluation Methodology

All Arena leaderboards use the Bradley-Terry (BT) pairwise comparison model, which infers a latent skill coefficient for each model from win/loss outcomes. LMArena's Arena-Rank Python package, open-sourced under Apache 2.0 and published on GitHub and PyPI, implements the BT fitting with closed-form confidence-interval calculation and a 30x speedup over the historical FastChat implementation. The package also supports reweighting to correct for non-uniform sampling, meaning models with fewer battles are not penalized. Style-control extensions layer additional covariates (token length, markdown header count, markdown bold count, markdown list count) into the BT regression to separate substance from stylistic formatting effects; coefficient estimation shows that length is the dominant style factor. For the Agent Arena, ranking uses causal tracing rather than pairwise votes: the arena treats each component selection as a treatment in a multi-intervention randomized controlled trial, aggregating multiple behavioral signals (confirmed success, praise/complaint, steerability, bash-error recovery, tool hallucination) into a single net-improvement estimate per model. The Arena-Hard pipeline (BenchBuilder, arXiv:2406.11939) automates benchmark construction by extracting hard prompts from the live Arena dataset via a seven-criterion hardness labeler; Arena-Hard-Auto v0.1 achieves 98.6% agreement with human preference rankings at a cost of approximately $20. The Search Arena methodology, published as a dataset and paper accepted at ICLR 2026 (arXiv:2506.05334), extends the BT model to search-augmented LLM systems across 11+ models from Perplexity, Gemini, and OpenAI. Prior to sharing any conversation data, LMArena uses GCP's Sensitive Data Protection API to remove personally identifiable information. Since July 2025, methodology updates are publicly logged in the Leaderboard Changelog. [CE003, CE009, CE013, CE014, CE015, CE016]

5.3 Deployment, Data Infrastructure, and Developer Ecosystem

The Arena-Rank ranking engine is published as a pip-installable Python package (pip install arena-rank) under the lmarena GitHub organization, which also hosts the list repository. The FastChat repository (lm-sys/FastChat, now primarily in maintenance mode) originally powered Chatbot Arena and remains a reference for the Vicuna model weights. The lmarena-ai HuggingFace organization hosts p2l (prompt-to-leaderboard) model variants used internally for preference modeling, and releases public battle datasets (e.g., arena-human-preference-140k) to support external research. As of June 2026 the changelog records ongoing model additions at a pace of several new models per week across leaderboards. The Code Arena's persistent sessions are built on Cloudflare R2 storage with CodeMirror 6 for source views and live rendering of generated web applications. Agent Mode sessions averaged approximately 16.5 structured tool calls per session, with roughly 75.6% of sessions using at least one tool; the highest-use sessions ran very long chains in a single week, covering coding, file creation, and web synthesis tasks. In one measured 7-day window, Agent Mode wrote 40.3 million lines of code across successful write_file calls, approximately 1,000 lines per coding session. Battles in Direct, integrated since May 2026, converts 10% of direct chat sessions into pairwise battles, with position-bias and same-org-indicator corrections applied to the BT fit. No API is publicly documented for third-party consumption of Arena scores or raw vote data, though open datasets are released periodically on HuggingFace. [CE013, CE014, CE015, CE019, CE030, CE031]

5.4 Differentiation, IP, and Research Advantage

LMArena's core differentiation is the combination of real-world, in-the-wild data at scale with methodological transparency. Unlike static academic benchmarks (MMLU, HumanEval) or automated LLM-as-judge pipelines, Arena captures genuine multi-turn user interactions across diverse languages, topics, and skill levels. Published research validates this approach: the original Chatbot Arena paper (arXiv:2403.04132) showed >80% agreement with expert raters; MT-Bench (NeurIPS 2023) demonstrated that LLM-as-a-judge can achieve human-level inter-rater agreement. The Arena-Hard benchmark (arXiv:2406.11939) achieves 3× higher model separation than MT-Bench at a fraction of the cost. The Search Arena ICLR 2026 paper (arXiv:2506.05334) extends the methodology to retrieval-augmented systems, demonstrating preference correlations with citation count and cited source types. LMSYS (the precursor nonprofit) published the foundational evaluation methodology papers that underpin the current company's differentiation; those papers are widely cited across the AI industry. The HuggingFace organization releases preference datasets (140k+ labeled battle pairs) enabling external researchers to audit rankings independently. The company's 9 published papers and 15+ blog posts covering evaluation methodology constitute a recognized body of IP. However, methodological assets are openly shared, which means rivals can replicate the algorithm while the company retains the data-network advantage (scale of battles) and community trust as the lasting moat. [CE003, CE016, CE035, CE036, CE037, CE039]

5.5 Trust, Safety, Quality Controls, and Adverse Evidence

LMArena's public leaderboard policy (last updated April 30, 2026) specifies eligibility criteria for model listing, a sampling policy requiring ≥20% of battles to be between publicly available models, a pre-release testing protocol for anonymized models, and a data-sharing policy using GCP's Sensitive Data Protection API to sanitize conversations before release. However, as of June 2026 LMArena has not published SOC 2 Type II or ISO 27001 certifications, no third-party security audit is publicly disclosed, and the help.arena.ai privacy policy page lacks enterprise data processing agreement (DPA) details. The platform has faced two serious credibility challenges. First, in April 2025 Meta submitted an "experimental, chat-optimized" version of Llama 4 Maverick that ranked #2 on the arena but was not the publicly released model; LMArena subsequently updated its policy and rescored the public version, which fell to 32nd place. Second, a paper authored by Cohere, Stanford, MIT, and AI2 in April 2025 alleged that certain large labs (Meta, OpenAI, Google, Amazon) received disproportionately high sampling rates and could selectively suppress low-performing pre-release variants, constituting benchmark gaming; LMArena denied the inaccuracies and pointed to its published sampling transparency. These incidents create ongoing reputational risk and have prompted calls from the research community for tighter pre-release limits, independent audits, and algorithmic transparency. The AI Evaluations commercial product offers SLA commitments but detailed uptime SLA terms and security disclosures are not public. [CE020, CE021, CE022, CE023, CE024, CE025]

5.6 Exhibits

6.1 Customer Base Segmentation

LMArena's customer base has two fundamentally different segments that must be distinguished carefully. The first is the free community, a global population of 5 million or more monthly users across 150+ countries who interact with the platform at no charge through arena.ai. These users submit prompts, vote on model outputs, and power the collective preference signal that underpins the leaderboards. This community spans developers, researchers, students, knowledge workers, and enthusiasts; per the two-year anniversary blog (April 2025), approximately 41% of battles involve open-source models, suggesting a technically sophisticated user base. The company generates no direct revenue from this segment but derives its core IP (preference data) and brand credibility from it. The second is the commercial customer segment, which consists of AI labs and enterprises that pay for the AI Evaluations service launched in September 2025. Named paying customers confirmed in the January 2026 Series A press release include OpenAI, Google, and xAI. Anthropic and Meta are major model providers on the leaderboard but their paying status as AI Evaluations subscribers is not separately confirmed in public disclosures; the TechCrunch 2026 article notes LMArena "partnered with" OpenAI, Google, and Anthropic for model submissions, a category that encompasses both commercial and non-commercial relationships. Enterprises (non-lab companies using AI Evaluations to benchmark models for their own applications) are a third stated target segment; no named enterprise non-lab customers have been publicly disclosed. [CU001, CU002, CU003, CU004, CU005, CU006]

6.2 Adoption Trajectory and Usage Metrics

LMArena's community adoption has been rapid and well-documented. The platform launched in May 2023 as a UC Berkeley research project; by May 2025 it disclosed the $100M seed round at a $600M valuation with significant traction. By January 2026, at the time of the Series A announcement, LMArena reported 5M+ monthly users (up from an implied ~3M cited in the September 2025 AI Evaluations blog), 60M+ monthly conversations, 250M+ cumulative conversations, and 2M+ monthly votes. The Series A blog post quantified community growth as "25x" since the seed round in May 2025. In terms of model evaluations, LMArena has evaluated 400+ public models and 300+ pre-release model variants. From September 2025 (commercial launch) to December 2025 (less than four months), the annualized revenue consumption run-rate reached $30M, per CEO Anastasios Angelopoulos and confirmed in TechCrunch's January 2026 article and the official press release. The Series A post also reported 50M+ community votes collected, 145k+ open-source battle data points released, and 400+ model evaluations across modalities. Notably, the platform's growth appears primarily pull-driven: the TechCrunch podcast episode notes that LMArena leaderboards "became something of an obsession among model makers," implying low customer-acquisition cost for the community segment. [CU001, CU002, CU003, CU004, CU007, CU008]

6.3 Named Customer Proof and Production Evidence

Named paying customer evidence for the AI Evaluations commercial product is limited but specific. The January 2026 PR Newswire press release states: "LMArena works with leading AI labs and enterprises, including OpenAI, Google and xAI, all drawing on LMArena's evaluations to improve their models for production use cases." This constitutes direct company-stated customer naming and is the strongest public customer-proof available. Separately, the Felicis partner Peter Deng stated in the same press release: "We're leading this round because LMArena has built the most trusted, reliable, real-world signal of AI performance. They have become essential infrastructure for every lab and enterprise." The company's Agent Arena methodology blog post documents real user sessions — a live sports-TV schedule site, a self-hosted movie watchlist, an autonomous underwater vehicle autopilot — as examples of production-style deployments, though these are community users, not paying enterprise customers. Prior to commercialization, Google's Kaggle, a16z, and Together AI had donated compute and grants to LMSYS (the precursor nonprofit). The lmsys.org historical launch records show the original 2023 partnership structure with UC Berkeley. For model providers that are not confirmed as paying: Anthropic and Meta are frequently cited as having their flagship models on the leaderboard, but the TechCrunch 2025 and 2026 articles consistently name only OpenAI, Google, and Anthropic as "partners" and only OpenAI, Google, and xAI as named commercial customers in the press release. No government procurement records, G2/Capterra reviews, or independent third-party case studies are publicly available for the AI Evaluations service. [CU005, CU006, CU012, CU013, CU014, CU015]

6.4 Retention, Durability, and Repeat Usage

LMArena has not publicly disclosed formal retention metrics such as net revenue retention (NRR), gross revenue retention (GRR), or customer churn for its AI Evaluations product. The platform is too young commercially (launched September 2025) to have multi-year cohort data. For the community segment, the principal retention signal is platform activity: the 25x community growth year-over-year (May 2025 to January 2026 per the Series A blog) and 250M+ cumulative conversations versus 60M/month active generation suggest that most of the cumulative usage reflects a growing ongoing community rather than one-time visits. The AI Evaluations product exhibits a consumption-based model ("annualized consumption rate" rather than subscription ARR), which means that retention depends on labs and enterprises continuously submitting evaluation requests. The rapid ramp from $0 to $30M annualized in under four months implies initial strong lab demand; whether that sustains as labs develop internal evaluation capabilities is a key risk. The leaderboard changelog shows continuous model additions (several per week in June 2026), which serves as an indirect proxy for ongoing model-provider engagement. The lack of any disclosed customer logo wall, case study site, or testimonial page (other than investor quotes from Felicis and UC Investments) is notable for a company at $1.7B valuation and $30M ARR run-rate. [CU007, CU009, CU010, CU011, CU016, CU017]

6.5 Adverse Evidence: Conflicts, Gaming, and Concentration Risk

LMArena's customer and model-provider relationships create well-documented structural risks. First, the conflict of interest: the same AI labs (OpenAI, Google, Anthropic, and others) that pay for AI Evaluations and serve as commercial customers are also the entities whose models are ranked on the public leaderboard. This creates an inherent incentive for labs to optimize for Arena performance, which they can do via the pre-release testing protocol. The TechCrunch podcast episode directly raises this: "how a team like theirs can build a neutral benchmark when the companies they're ranking are also their backers." Second, the Meta Llama 4 Maverick incident (April 2025): Meta submitted a "chat-optimized experimental" version of Maverick to the arena that achieved a #2 ranking, but the publicly released model ranked approximately 32nd. LMArena updated its policy and stated "Meta's interpretation of our policy did not match what we expect from model providers." This incident shows that the commercial partnership structure (which depends on labs submitting models) creates practical pressure to accommodate behavior that damages leaderboard integrity. Third, the Cohere/Stanford/MIT/AI2 study (April 2025) alleged that Meta could privately test 27 model variants between January and March 2025 before selecting the highest performer, that sampling rates were disproportionate for top labs, and that the study's preliminary findings were not disputed by LMArena when shared. LMArena denied the specific claims but announced a new sampling algorithm. These allegations have not been independently adjudicated. Fourth, customer concentration: with only three publicly named paying customers (OpenAI, Google, xAI), LMArena's commercial revenue is highly concentrated; if any of these labs develop or contract an alternative evaluation vendor, the impact on ARR could be significant. [CU005, CU006, CU012, CU019, CU020, CU021]

6.6 Exhibits

7.1 Conflict of Interest and Benchmark Integrity Risk

LMArena's cardinal risk is the structural conflict between its independent-arbiter identity and its revenue model. The company charges the same AI laboratories — OpenAI, Google, xAI — whose models appear on its public leaderboard for paid evaluation services. This dual role creates incentives, real or perceived, for preferential treatment. Academic researchers from Cohere, Stanford, MIT, and AI2 documented these dynamics in "The Leaderboard Illusion" (arXiv 2504.20879), finding that a handful of large providers could privately test multiple model variants, selectively disclose only their best scores, and receive disproportionately more evaluation battles. Meta's Llama 4 Maverick episode — in which Meta submitted a conversationality-optimized model to LMArena that it never publicly released, achieving a top-two ranking before the discrepancy was publicly exposed — demonstrated that policy gaps can allow benchmark gaming even without explicit rule violations. LMArena updated its policies in April 2026 but the fundamental tension between commercial dependency and perceived neutrality persists. Commentators including independent researcher Gwern have called the leaderboard "a cancer" and questioned whether its signal retains scientific value. The ucstrategies.com analysis found that models tuned specifically for Arena preferences can inflate scores by up to 100 Elo points, while SurgeAI analysis found evaluators disagreed with LMArena votes 52% of the time. If the credibility narrative fractures permanently — through a high-profile bias investigation, a paying-lab scandal, or a regulatory inquiry into AI benchmark accuracy claims — LMArena's entire product value proposition collapses simultaneously in both its consumer-facing leaderboard and its enterprise evaluation revenue. [CR001, CR002, CR003, CR004, CR005, CR006]

7.2 Regulatory, Legal, and Data Privacy Risks

LMArena operates under a complex and rapidly evolving regulatory environment. The legal entity, Arena Intelligence, Inc. d/b/a LMArena, processes personal data from 5 million monthly users across 150 countries, triggering obligations under the EU General Data Protection Regulation (GDPR), the EU AI Act (general-purpose AI model obligations active since August 2025), the Digital Services Act (potentially VLOP thresholds at scale), and a patchwork of US state privacy laws including California's CPRA. LMArena's own privacy policy (effective September 2025) explicitly warns users that prompts and generated responses may be shared publicly, raising user-consent adequacy risks under GDPR Article 7. The right-to-erasure challenge for trained evaluation models (GDPR Article 17) is a known compliance gap across the AI industry with no settled enforcement interpretation. The EU AI Act's general-purpose AI documentation and transparency requirements impose additional operational overhead. LMArena's practice of sharing user prompts with AI providers for model improvement purposes creates data-flow obligations that must be covered by data processing agreements in each jurisdiction. The EU AI Act's definition of "high-risk" AI system has not explicitly named AI benchmarking platforms, but evaluations in regulated domains (law, medicine, employment) could attract sector-specific scrutiny. No litigation directly involving LMArena or Arena Intelligence, Inc. was found in public records as of the run date; however, the company has not disclosed its compliance attestation status or any regulatory inquiries. IP risk from user-submitted prompts used in training is an unsettled area of law across multiple jurisdictions. [CR007, CR008, CR009, CR010, CR011]

7.3 Operational, Technical, and Security Risks

LMArena's operational risks center on platform reliability, data quality integrity, and the security of its community-generated evaluation dataset. With 60 million monthly conversations flowing through its infrastructure, a sustained outage or data breach would directly damage the real-time, continuous-evaluation proposition that differentiates LMArena from static benchmarks. The platform depends heavily on cloud infrastructure (compute costs for running multiple live AI model APIs simultaneously) whose pricing and availability are controlled by third parties including the same AI labs LMArena evaluates. A provider pulling API access — feasible if a lab disputes a rating or competitive dynamic shifts — would immediately create gaps in the leaderboard. Statistical integrity is another operational risk: LMArena's Elo/Bradley-Terry scoring relies on a sufficiently uniform and unbiased distribution of battles; any systematic prompt injection, coordinated voting campaigns, or sybil attacks on the platform's user base would corrupt the ranking signal without necessarily being detectable in real time. The shift to "Battles in Direct" (May 2026) where 10% of direct-chat sessions generate battle votes required post-hoc bias corrections for position bias and same-organization indicator bias, demonstrating that methodological updates can shift model rankings after the fact. Security of the evaluation dataset — a commercially valuable asset — against model distillation attacks or unauthorized scraping is an ongoing concern, particularly given that leading AI labs have financial incentives to study the evaluation distribution. Employee headcount of ~41 creates key-person concentration for a platform of this scale. [CR012, CR013, CR014, CR015, CR016]

7.4 Partner Concentration, Financial, and Execution Risks

LMArena's commercial model depends on a small set of large AI labs and enterprises for the majority of its $30M annualized revenue. With approximately 100 paying customers as of early 2026 and the top clients being OpenAI, Google, and xAI, revenue concentration is material: the loss of even one or two Tier-1 lab relationships could represent a disproportionate revenue decline. The same labs are simultaneously LMArena's best customers and its most motivated potential gaming adversaries, creating a tension with no clean resolution. The $1.7 billion valuation implies a ~57x revenue multiple on $30M ARR, pricing in substantial growth expectations that require maintaining trust, expanding into new domains, and upselling existing customers. At 41 employees, execution capacity is thin relative to the product surface area LMArena has already committed to (Agent Arena, WebDev Arena, Vision, Video, Search, Code, Document leaderboards plus commercial evaluations). The company's origins as a UC Berkeley research project mean that founding team academic ties and publication obligations may divert leadership attention from commercial execution. Cash burn dynamics are undisclosed; at $250M raised and 41 headcount, the runway appears long, but infrastructure costs at 60M monthly conversations are substantial. Future funding risk is moderate given market conditions but not zero: if the AI evaluation market fails to grow to the $3.8B projected by 2030, subsequent rounds may price in disappointment. The $382,500 Recall Capital-LMArena SEC Form D filing (February 2026) reflects secondary-market investor demand but is not evidence of the company's own financial health. [CR017, CR018, CR019, CR020, CR021]

7.5 Exhibits

8.1 Financing and Valuation Context

Arena Intelligence, Inc. (d/b/a LMArena) raised $150 million in a Series A round announced January 2026, achieving a post-money valuation of $1.7 billion. This followed a $100 million seed round at a $600 million valuation in May 2025, bringing total capital raised to $250 million in approximately seven months. The Series A was led by Felicis and UC Investments (University of California), with participation from Andreessen Horowitz, The House Fund, LDVP, Kleiner Perkins, Lightspeed Venture Partners, and Laude Ventures. The company's annualized "consumption run rate" — which it describes as equivalent to ARR — surpassed $30 million in December 2025, less than four months after launching its first commercial product (AI Evaluations) in September 2025. This gives a post-money valuation-to-ARR multiple of approximately 57x on the $30M run rate, compared with 8-15x typically seen for late Series A SaaS companies. However, the trajectory from $0 ARR (May 2025) to $30M ARR (December 2025) in four months is exceptional and investors are pricing forward growth, not current revenue. The SEC EDGAR Form D (accession 0002113470-26-000001, filed February 2026) for the "Recall Capital-LMArena" secondary fund at $382,500 indicates secondary-market demand at implied valuations consistent with the Series A price. Cap table details — dilution from seed to Series A (approximately 17% sold at seed, approximately 9% sold at Series A) imply a pre-money Series A enterprise value of $1.55 billion based on disclosed post-money valuation and round size. No public evidence of convertible notes, preferred liquidation preferences, or prior down-round risk has been found; the company raised at a 2.83x step-up in eight months. [CV001, CV002, CV003, CV004, CV005, CV041]

8.2 Valuation Analysis — Multiple Lenses

Revenue-multiple analysis: At $30M ARR and $1.7B valuation, the ARR multiple is 57x — placing LMArena in the top decile of late-stage AI infrastructure funding rounds. The comparable universe for AI evaluation infrastructure is thin: Weights and Biases was acquired by CoreWeave for $1.4 billion in March 2025 (a training/evaluation platform with broader tooling), Scale AI has been valued in the multi-billion range with revenues an order of magnitude larger. For early-stage pure-play AI benchmarking, no direct public comparables exist. The most relevant analogies are AI "picks and shovels" infrastructure: LMArena's position as a neutral data layer resembles how Bloomberg or S&P Global serves financial markets — a trusted data franchise with high switching costs. The AI model evaluation market was $1.86B in 2025 and is projected to reach $2.36B in 2026 and $6.24B by 2030 at 27.3% CAGR per The Business Research Company. If LMArena captures 5-10% of a $3B market by 2028, revenue of $150-300M at a 15-20x multiple implies a $2.25-6B valuation — the base-to-bull range for a public-comparable exit. Growth sustainability: reaching $30M in four months implies an implied growth rate of $7.5M per month. Sustaining even 50% of this adds $45M ARR in the next 12 months, reaching $75M ARR by end-2026. At a compressed 25x multiple (consistent with a maturing enterprise SaaS franchise), that implies a $1.875B valuation — roughly flat to current. At 20x multiple (public SaaS median for high-growth), $75M ARR yields $1.5B. The current $1.7B valuation thus depends on either (a) sustaining rapid ARR growth, (b) maintaining the premium platform multiple, or (c) both. Scenario analysis below reveals the distribution of outcomes more explicitly. [CV006, CV007, CV008, CV009, CV010]

8.3 Investment Thesis and Anti-Thesis

The investment thesis rests on three durable structural advantages: (1) community network effects — 5M monthly users generating 60M conversations create a data moat that is extremely difficult to replicate from scratch even with significant capital; (2) incumbency in a critical judgment role — leaderboard rankings influence AI lab marketing, developer adoption, and enterprise procurement in a way that creates switching costs for LMArena's paying customers; (3) market timing — real-world human preference data for AI evaluation will become more valuable, not less, as the number of competing frontier models proliferates and enterprise buyers face genuine selection complexity. The anti-thesis centers on four concerns: (1) the structural conflict of interest between evaluation revenue and evaluated subjects (documented by the Leaderboard Illusion paper) could precipitate a trust collapse that simultaneously destroys both the public leaderboard and the enterprise revenue; (2) Goodhart's Law dynamics — as Arena becomes the de facto standard, labs optimize specifically for it, degrading the signal's real-world relevance; (3) the revenue multiple is extreme and prices in a growth trajectory that requires both sustained execution and continued credibility; (4) a platform dependency risk exists where the very AI labs that are paying customers control the API access that makes the platform function, creating a structural leverage imbalance. On balance, the thesis is stronger than the anti-thesis at the right entry price, but the current 57x ARR multiple leaves insufficient margin of safety given the unresolved conflicts. [CV011, CV012, CV013, CV014]

8.4 Recommendation, Scenarios, and Final Diligence Asks

The overall recommendation is "track" with a path to "buy" conditional on: (a) receipt of an independent methodology audit responding to arXiv 2504.20879 findings; (b) entry at a valuation below 30x next-twelve-months ARR (implying ARR of at least $57M before committing at the current $1.7B price); and (c) commercial contract terms confirming no preferential evaluation pricing for paying lab customers versus public leaderboard methodology. Risk rating is "high" given the conflict-of-interest and multiple-compression risks. The confidence level in this recommendation is "medium" — sufficient evidence exists to form a view, but the key risk factors are structural and unresolved. Valuation stance is "stretched" at current price. The company's trajectory, community engagement, and first-mover advantages are genuine; the price paid for them is not. Bull case requires $75M+ ARR by end-2026 with maintained credibility; bear case is a credibility collapse in which both revenue and valuation deflate to $200-400M (similar to benchmark reputation crises in other industries). Investors with long hold periods and strong governance influence who enter at 30x forward ARR or below have a credible risk/reward scenario in the base case. [CV015, CV016, CV017, CV018, CV019]

8.5 Exhibits