By the time most startups figure out how to charge for their product, they've already burned through their seed capital and worn out their first pitch deck. Arena Intelligence skipped that part entirely.
Eight months after launching its commercial service, the UC Berkeley spinout hit $100 million in annualized revenue—a figure the company announced on June 29, 2026, that reflects usage-based consumption rather than locked-in subscriptions. But even with that caveat, the trajectory is remarkable. Arena went from zero commercial revenue in September 2025 to a $30 million annualized consumption run rate by December, then watched that number triple again by mid-year.
The velocity, according to CEO Anastasios Angelopoulos, stems from an uncomfortable reality for AI labs: they all need what Arena built, and the alternatives aren't good enough.
The Leaderboard Nobody Planned to Monetize
Arena didn't begin with a business model. It started as Chatbot Arena, a 2023 research project out of Berkeley's LMSYS and Sky Computing Lab. The concept was almost stupidly simple—let people compare AI models in blind matchups, aggregate the votes, publish rankings. Academic experiments rarely scale beyond a few thousand participants and a conference paper.
This one did. By May 2025, before Arena even formally incorporated, the leaderboards were drawing over a million monthly visitors. Somewhere along the way, the AI industry had decided Arena's crowdsourced rankings mattered more than any benchmark dreamed up in a lab. Model releases now came with an implicit question: where will it land on Arena?
Angelopoulos, a Berkeley postdoc who'd studied under Michael I. Jordan and Jitendra Malik, recognized the gap before most. Labs were lighting billions of dollars on fire for training runs, yet evaluation infrastructure remained a patchwork of unreliable tools and contaminated benchmarks. His co-founder, Wei-Lin Chiang—a Berkeley PhD student who'd helped build Vicuna, FastChat, and the original Chatbot Arena—had already laid the technical groundwork. Ion Stoica, the Berkeley professor behind Databricks and Anyscale, signed on as the third founder.
They incorporated in April 2025. One month later, they closed a $100 million seed round at roughly $600 million post-money, led by Andreessen Horowitz and UC Investments. By January 2026, the company had rebranded from LMArena to simply Arena and locked down a $150 million Series A at a $1.7 billion valuation. Total capital: $250 million, raised before most companies figure out product-market fit.
What Labs Actually Pay For
The commercial offering, dubbed AI Evaluations, launched September 16, 2025. It builds on the same crowdsourced foundation as the public leaderboards but wraps enterprise infrastructure around it: SLAs, custom evaluation frameworks, auditability. Labs can tap into Arena's dataset of real-world usage patterns—the messy, unpredictable ways people actually interact with AI models, not how researchers think they should.
At commercial launch, the platform had logged 250 million conversations and was processing 2 million votes monthly. Nine months later, those figures look almost quaint. As of late June 2026, Arena reported 10 million monthly visitors, 700 million cumulative conversations, and 82 million votes cast. The growth reflects both continued public obsession with the free leaderboards and enterprise adoption of the paid evaluation tools.
The customer list, according to Arena's January 2026 press release: OpenAI, Google, xAI. Labs racing to ship frontier models don't just compete on Arena's public rankings anymore. They pay for the machinery underneath.
In February 2026, Arena launched Max, a model router that uses over 5 million community votes to direct requests to different models based on task requirements. The company claimed Max variants topped its internal leaderboards at launch—a bit of self-referential marketing, perhaps, but also a demonstration of what's possible when you own both the evaluation data and the distribution infrastructure.
The Agent Bet
The potentially more revealing product move came in June 2026: Agent Mode. The feature lets users and evaluators test AI agents in sandboxed environments equipped with real tools—web search, image generation, coding environments, bash terminals. Within a month, Agent Mode was logging 5 million turns monthly, growing roughly 10 percent week-over-week.
Arena launched an accompanying Agent Arena leaderboard that ranks 28 orchestrator agents not on conversational fluency but on task completion, tool reliability, and resistance to hallucination. By late June, the agent leaderboard had processed over a million sessions. Early usage patterns showed coding accounted for 29 percent of activity, research 11 percent, planning another 11 percent.
It's a bet that the evaluation game is shifting from static chat responses to dynamic, multi-step workflows. Arena is positioning itself as the testing ground for that transition—and hoping enterprise customers will pay to evaluate agents the same way they now pay to evaluate chatbots.
Gaming the System, or Just Early Access?
The ascent hasn't been entirely smooth. In April and May 2025, outlets including Computerworld, TechCrunch, and Ars Technica reported concerns about what critics called the "leaderboard illusion." The allegation: major labs were pre-testing models privately on Arena before public release, potentially gaming the rankings to guarantee a strong debut.
Arena pushed back with data showing any advantage from early testing evaporated as fresh votes poured in. The company also pointed to ongoing policy changes and transparency measures. An ICML 2025 paper by multiple researchers explored adversarial manipulation risks, including vote-rigging and de-anonymization attacks, alongside proposed countermeasures.
But the broader critique cuts deeper. Multiple observers have questioned whether any human-preference leaderboard, Arena included, can capture model capability comprehensively. Static benchmarks like MMLU have known contamination issues—models trained on test data, essentially. Arena's live, in-the-wild approach arguably addresses some of those concerns, though questions about representativeness and potential bias linger.
Angelopoulos told TechCrunch in June 2026 that Arena now competes less with other leaderboards—Yupp shut down in March 2026—and more with traditional human-labeling firms like Mercor, Surge, and Scale AI for a slice of post-training budgets. That's a telling shift: Arena is no longer just a scorekeeper. It's selling picks and shovels for the AI gold rush.
Infrastructure Disguised as Research
Arena's mission statement reads: "measure and advance the frontier of AI for real-world use." The $100 million revenue milestone in eight months suggests the market agrees that mission is valuable, perhaps even essential.
The company now operates leaderboards across chat, code, image, video, and agents. It maintains open datasets and publishes academic papers. It runs an Academic Partnerships Program offering up to $50,000 per project for university researchers. It partnered with DataTecnica on BiomedArena, bringing NIH benchmarks into the platform.
Whether Arena becomes the enduring standard for AI evaluation or just a waypoint in a fast-moving market remains uncertain. Standards are funny things—they either calcify into infrastructure or get disrupted by the next wave of innovation. For now, Arena has captured something uncommon: a research artifact that morphed into infrastructure, and infrastructure that labs actually pay for.
Eight months to $100 million suggests they're paying quite a lot. The question is whether Arena can maintain that momentum as the labs themselves build internal evaluation capabilities, or whether it's already too deeply embedded in their workflows to extract. Angelopoulos is betting on the latter. So far, the numbers suggest he might be right.
