Arena — the platform previously known as Chatbot Arena and LMArena — is the closest thing the AI industry has to an independent rankings table. It collects millions of blind votes from real users and turns them into a leaderboard that AI companies track closely — and occasionally try to game.

How the Voting Works

The mechanism is simple. Visit arena.ai, type any question, and two anonymous AI models respond in parallel. You vote for whichever gave the better answer. Only after you vote does the platform reveal which models you just compared.

This blind structure is deliberate: it prevents users from unconsciously favouring a brand they already trust. The scores use a Bradley-Terry model — a statistical method that estimates each model’s true strength from the full matrix of head-to-head comparisons. Beating a top-ranked model earns more than beating a weaker one, and the leaderboard updates continuously as new votes arrive.

Why Companies Submit Models Anonymously

Arena lets AI labs submit unreleased or experimental models under an “Anonymous” label. Users vote on them without knowing the source, and results flow back to the provider privately. Only after the lab chooses to reveal the model — or after enough votes have stabilised its score — does the identity appear.

The practice serves two purposes. First, it gives labs objective data before a public launch: if the model underperforms, no reputation is at risk. Second, it removes the halo effect — a model labelled “GPT-something” or “Claude-something” might attract votes simply because the name is trusted, distorting the comparison.

A recent illustration: Meituan’s LongCat-2.0 topped global Arena charts for two months under an anonymous label before the company revealed its identity. Stealth testing on Arena has become a standard pre-launch tool for major AI labs.

What the Leaderboard Shows

Arena tracks more than 90 commercial and open-weight models across a general-purpose leaderboard and specialised tracks for coding, vision, web development, and search. As of mid-2026, reasoning-optimised models dominate the top tier, with models from Anthropic, Google, OpenAI, and xAI clustered closely at the top.

The leaderboard refreshes continuously and is publicly accessible — no account required to view rankings.

How to Use It Yourself

Arena is completely free and requires no registration. Visit arena.ai and choose a mode:

  • Battle mode — submit a prompt and vote between two anonymous models.
  • Direct comparison — choose specific models to compare head-to-head.
  • Specialised arenas — separate leaderboards for coding (Copilot Arena), visual reasoning, search, and more.

One important note: conversations may be recorded and used for AI research. Do not submit personal or sensitive information.

Limitations and Criticism

Arena is influential, but not without flaws.

Selective transparency. Labs can privately test many model variants and only publicise the winner’s result — showing only their best face to the public record.

Goodhart’s Law. When a benchmark becomes a target, it ceases to be a reliable measure. AI researchers have observed that labs now optimise models specifically to score well on Arena, rather than treating it as a neutral proxy for real-world usefulness.

Preference ≠ quality. A fluent, confident-sounding response can outscore a more accurate but plainer answer. Human preference and factual correctness don’t always align.

Suspected bias. A 2025 paper argued that Arena’s voting patterns may systematically favour outputs from certain large labs — a claim Arena has disputed.

Arena remains the most widely referenced independent AI leaderboard. But like all benchmarks, it measures what it measures — not everything that matters.

In the News

When Meituan revealed that LongCat-2.0 had secretly led global AI leaderboards for two months under an anonymous label, it was Arena’s anonymous submission system at work — a case study in how labs actually use the platform before a major launch.

FAQ

Is Arena (Chatbot Arena) free to use?
Yes. No account, no API key, no payment required. Visit arena.ai and start voting immediately.

What happened to LMSYS Chatbot Arena and lmarena.ai?
The platform was launched by LMSYS Org at UC Berkeley in May 2023 and lived at chat.lmsys.org, then lmarena.ai. In January 2026 it spun out as an independent company and rebranded to Arena at arena.ai.

Can AI labs cheat the leaderboard?
Labs can optimise models for Arena-style tasks, inflating scores without improving general usefulness — a known concern. Submitting the same model repeatedly under different labels is against the rules, but testing many private variants and revealing only the best is permitted.

How many votes does a model need for a reliable ranking?
New models start with wide uncertainty intervals that tighten as comparisons accumulate. Thousands of head-to-head votes are needed for stable rankings. The platform has collected over 800,000 votes since its launch in 2023.