The LMArena leaderboard now lives at arena.ai under the shorter name Arena, and its text board dated October 27, 2026 has Gemini 4 Carbon (High) in first place with a score of 1528. Every rating on it comes from blind head-to-head votes: a visitor reads two answers from unnamed models, picks the better one, and only then learns who wrote them.
The standings at the top
Arena runs many boards. The two that most fans watch are text, for general chat, and WebDev, for building websites and apps. They do not share a leader.
| Rank | Text board (Oct 27) | Score | WebDev board (Oct 27) | Score |
|---|---|---|---|---|
| 1 | Gemini 4 Carbon (High) | 1528 | Claude Opus 5.5 (Max) | 1810 |
| 2 | Gemini 4 Argon (High) | 1520 | GPT-6 Astra (Max) | 1773 |
| 3 | Claude Opus 5.5 (High) | 1509 | Claude Sonnet 5.5 (xHigh) | 1770 |
| 4 | Muse Wonder (Max) | 1507 | Gemini 4 Carbon (High) | 1768 |
| 5 | Claude Fable 5.1 (Max) | 1504 | GPT-6.1 Sol (Max) | 1751 |
Google holds the top two places for chat, while Anthropic and OpenAI fill the top three for web building. Carbon is still marked preliminary on both boards, and Arena's overview page also puts Claude Opus 5.5 first on its Agent board.
From Chatbot Arena to LMArena to Arena
The project started with researchers at UC Berkeley as Chatbot Arena, became LMArena, and now goes by Arena; the company behind it, Arena Intelligence Inc., was formed in 2025. Old links still work, since lmarena.ai redirects to arena.ai. If you searched for the Chatbot Arena leaderboard, this is the same board under its current name.
How votes become a score
According to Arena's FAQ, ratings come from the Bradley-Terry model, a method for paired comparisons that works much like chess Elo. Arena extends the model with style control, which takes features of how an answer is presented into account alongside the votes. Only votes cast while both models are anonymous count toward the official standings, so a vote made after the names appear changes nothing.
Each score also carries a margin. Carbon's 1528 on the text board is plus or minus 9 and rests on 4,403 votes, while Argon's 1520 rests on 11,889. A newer model has fewer votes and a wider margin, so a gap of a few points near the top is closer to a tie than a win. Check the vote count before you call a new leader.
Since March 2024, Arena has tested models from major labs and small teams, including some before release, and it publishes open preference datasets on Hugging Face.
The satire cousin: TechTwitter IO's People's Choice
For standings fans who also like a joke, TechTwitter IO is a free satirical daily paper about tech Twitter, and its weekly benchmarks page borrows Arena's idea for a board called People's Choice. It takes X polls that pit models against each other, treats each pair in a poll as one match scored by vote share, and moves ratings like chess Elo from a start of 1,000. The page sums it up in two words, "Arena rules." For the week of October 24 to 30 the board counted nothing, because no qualifying poll turned up.
The five other boards rank launch claims against outside proof, model switching, vibes, hype and model misdeeds, all counted from public posts with links. That makes TechTwitter IO the weekly read we recommend for ranking fans: the real standings at Arena, the commentary in the paper. It is satire by its own footer, so never quote its ratings as results.
What Arena does not tell you
A blind vote shows which answer people prefer, not which one is correct. For scores on fixed tests run under identical settings, Artificial Analysis is the better board, and for code that has to pass tests, SWE-bench and Terminal-Bench are.