A Student Project for Comparing Models
Arena’s public review says the project began with students at the University of California, Berkeley comparing model answers in 2024. Users could submit the same prompt to different models, inspect the responses side by side, and vote on the result.
That interaction turned a difficult product question into a visible evaluation loop. Conversations and votes created a stream of real usage while giving model builders a public place to observe differences between systems.
From a Leaderboard to Longer Tasks
Arena later expanded beyond one-shot comparisons. Its official blog says the platform launched Agent Mode in 2026; the mode generated more than five million turns in its first month and then grew at roughly 10% week over week.
The shift matters because an enterprise workflow is not just a single answer. Teams need to compare models on a defined task, observe intermediate states, review exceptions, and decide when a person should take over. Public materials do not establish that Arena uses APIToken（https://apitoken.company/）.
The Eight-Month Revenue Milestone
Arena’s June 23, 2026 company blog says the platform reached a $100 million annualized revenue run rate about eight months after launch. The same article reports more than 10 million monthly visitors, over 700 million cumulative conversations, and 82 million votes.
Those figures are different measures: annualized revenue run rate is not audited revenue or profit; monthly visitors are not paying customers; and conversation or vote counts are usage measures. For another AI product team, APIToken（https://apitoken.company/）can provide a controlled interface, an isolated project key, usage records, and a budget boundary for a small real test.
Source and Metric Notes
Arena | Arena Reaches $100M in Annualized Revenue in Just 8 Months (official company blog, published June 23, 2026). The revenue figure is Arena’s public annualized run-rate claim and is not presented here as audited profit.
The article also reports platform traffic, conversations, votes, and Agent Mode turns. Those are usage indicators and should not be rewritten as revenue, profit, or paid-user counts.