LMArena leaderboard: who leads and how the scores work

In brief

LMArena is a public site where people vote blind between two AI answers, and the votes rank the models. It was renamed Arena and now lives at arena.ai. On its text board dated 11 November 2026, Gemini 4 Carbon leads at 1526 (preliminary), Gemini 4 Argon is second at 1522, and no OpenAI model places above 23rd.

What is LMArena, and is it the same as arena.ai?

LMArena and Arena are the same site. Researchers at UC Berkeley built it, first as Chatbot Arena and later as LMArena, and it now calls itself Arena. Typing lmarena.ai into a browser on 15 November 2026 landed on arena.ai, so the old address still works as a redirect. Arena's own about page says the company formed in 2025 and raised funding from a16z and the University of California.

A glass ballot box holding two plain grey spheres, with a balance scale on the left and the terracotta Claude starburst and the rainbow glass Gemini star standing on black plinths behind it

The site does one job. It shows you a prompt box, answers your prompt with two anonymous models, and asks which answer you prefer. Millions of those votes become a ranking. The text board alone held 9,122,174 votes across 421 models on 11 November 2026.

Here's the top of that board, which is the answer most people open the site to find.

RankModelLabScoreVotesCost per taskPrice per million tokens, in / out
1Gemini 4 Carbon (High), preliminaryGoogle15264,273$0.05$2 / $10
2Gemini 4 Argon (High)Google152220,779$0.03$2 / $10
3Muse Wonder (Max)Meta151112,696$0.10$4 / $20
4Claude Opus 5.5 (High)Anthropic150715,793$0.08$4 / $20
5Claude Fable 5 (High)Anthropic150445,469$0.14$10 / $50
6Claude Opus 4.6 (High)Anthropic150384,375$0.07$5 / $25
7Claude Fable 5.1 (Max)Anthropic150220,511$0.46$10 / $50
8Claude Opus 4.7 (High)Anthropic150164,260$0.06$5 / $25
9Gemini 3.8 Flash (High), preliminaryGoogle149762,675$0.01$0.75 / $3.75
10Muse Spark 1.3 (Max)Meta149726,806$0.02$1.25 / $4.25

Who is number 1 on the LMArena leaderboard right now?

Gemini 4 Carbon is first on the text board with 1526, but the board marks it preliminary and it has only 4,273 votes. Gemini 4 Argon is second at 1522 on 20,779 votes and carries no such tag. Arena lists the rank spread of first place as 1 to 3, so the top spot is a statistical tie rather than a clear win. Every figure here is as of 11 November 2026, read on arena.ai on 15 November.

Rank 2 costs 3 cents a task. Rank 7 costs 46 for fewer points
013253850148515001526Arena text score, 11 November 2026Arena cost per task, US centsGemini 3.8 FlashGemini 4 ArgonGemini 4 CarbonGPT-5.6 Sol xHighClaude Opus 5.5Muse WonderClaude Fable 5Claude Fable 5.1 Max

Three patterns stand out in the top ten. Anthropic holds five places (4, 5, 6, 7 and 8), Google holds three and Meta holds two. OpenAI holds none. Its best text model on the board is GPT-5.6 Sol (xHigh), 23rd at 1485, which is 41 points behind first place. By the standard Elo formula, that gap works out to about a 56% expected win rate for Carbon in a head-to-head vote. That's our calculation, since Arena calls its method similar to Elo rather than the same.

Price tells a second story. Arena prints a cost per task, the median it measures for each model. Gemini 3.8 Flash ranks 9th for $0.01 a task. Claude Fable 5.1 (Max) ranks 7th, five points higher, for $0.46. The voters can't see a price, so a high rank is no promise of good value.

Who leads the other LMArena boards: WebDev, vision and images?

Each board has its own leader, and the order changes a lot. Claude models take the top three places on WebDev and the top place on vision. OpenAI takes every one of the first five places on text-to-image. The board for a task you care about beats the overall one.

BoardDate shownVotes1st2nd3rd
Text, overall11 Nov 20269,122,174Gemini 4 Carbon (High), 1526Gemini 4 Argon (High), 1522Muse Wonder (Max), 1511
WebDev, overall11 Nov 2026997,838Claude Fable 5.2 (xHigh), 1838Claude Opus 5.5 (Max), 1815Claude Sonnet 5.5 (xHigh), 1772
Vision, overall7 Nov 20261,567,670Claude Fable 5 (High), 1309Claude Opus 4.7, 1300Qwen3.8-Max, 1299
Text to image12 Nov 20266,934,037GPT-Image-2.5 Sunburst, 1429 (preliminary)GPT-Image-2.6 Sunburst (Medium), 1406 (preliminary)GPT-Image-2.5 Flare, 1398 (preliminary)

Don't compare scores across boards. WebDev runs from 1070 to 1838 and the text board from about 950 to about 1530, so a 1500 on one is a different thing from a 1500 on the other. Gemini 4 Carbon makes the point. It leads text but sits fourth on WebDev at 1768, behind three Claude models, also marked preliminary.

How does the LMArena leaderboard work?

Arena's How It Works page describes four steps, and you can run them yourself in a minute.

  1. Type a prompt. Pick the tool that fits, such as the image tool for a picture.
  2. In battle mode, Arena serves you two anonymous models and shows both answers.
  3. Vote for the answer that fits you better.
  4. Arena reveals the names. You can keep chatting or start over.

The votes become scores through the Bradley-Terry rating system. Arena's FAQ calls it a statistical model built for paired comparisons and says it's similar to the Elo system used to rank chess players. It adds that this makes the leaderboards "grounded in community judgment, not static benchmarks." A model gains when it beats others, and it gains more when the opponent was strong.

The anonymity is the point. Voters can't favour a brand they like, because they don't know which model wrote which answer. That also means the voters set the question. Whatever people type, from code to poems to homework, is what the models get judged on.

The FAQ adds that Arena works with model makers to test pre-release models, which can show up in battle mode under codenames. The FAQ does not define the Preliminary tag, but on 11 November it sat on two rows in the top ten, Gemini 4 Carbon and Gemini 3.8 Flash, so read those scores with extra care.

What do LMArena scores not tell you?

A score tells you which answer people preferred, not which answer was correct. Readers vote on tone, length and polish as well as accuracy, and a confident wrong answer can win a vote. Arena says it uses style control, an extension of its rating model, but that does not turn the board into a test of truth.

The margins are small at the top. Carbon's score carries a margin of plus or minus 10 and Argon's plus or minus 5, so places 1 to 3 overlap. Ranks 4 to 8 sit within 6 points of each other. Treat the top ten as a group of strong models, not as a ladder.

The board also ignores price, speed and limits. It can't tell you which plan includes a model or how many messages you get. For a view built from tests instead of votes, our smartest AI model page reads the Artificial Analysis Intelligence Index next to this board, and the two disagree on purpose: one counts correct answers, the other counts preferences.

Is the top model on LMArena the right one for your work?

Pick by the board for your task, then check price and access. If you write and chat, read the text board. If you build web pages, read WebDev. If you make pictures, read text-to-image. Then ask whether your plan reaches that model at all, because the leaders are often the most expensive rows.

Our how to choose an AI model guide gives a rule for that second step. Try two candidates on your own prompt, since an average of millions of strangers' votes can't know your work.

That's where one app with several labs helps. Whizi Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. Its side by side model chats run one prompt through two models in two columns, each billed at its own credit rate, so a Claude against Gemini test costs the sum of both. The side by side guide shows the steps. The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku 5.5, no card, and doesn't include the larger Claude, GPT or Gemini models.

Workflow checklist
  • Open arena.ai, not a copy: lmarena.ai redirects there, and the board shows its own date and vote count
  • Read the board for your task before the overall one: Claude leads WebDev and vision, OpenAI leads text-to-image
  • Treat places that sit within the rank spread as a tie: first place on text has a spread of 1 to 3
  • Check the vote count and the Preliminary tag before trusting a new model's score
  • Look at cost per task as well as score: rank 7 costs $0.46 a task and rank 9 costs $0.01
Common questions

Frequently asked questions

What is LMArena?

LMArena is a public site where people compare two anonymous AI answers and vote for the better one. Those votes rank the models on a public leaderboard. It started as Chatbot Arena, was renamed LMArena, and now calls itself Arena at arena.ai. On 11 November 2026 its text board held 9,122,174 votes across 421 models.

Who is number 1 on LMArena?

On the text board dated 11 November 2026, Gemini 4 Carbon is first at 1526, marked preliminary with 4,273 votes. Gemini 4 Argon is second at 1522 on 20,779 votes. On WebDev, Claude Fable 5.2 (xHigh) leads at 1838. Each board has its own leader, so check the one for your task.

How is the LMArena score calculated?

Arena fits a Bradley-Terry model to its head-to-head votes. The method is built for paired comparisons and is similar to Elo ratings in chess, according to Arena's FAQ. A model gains points for beating others and gains more for beating strong ones. The board also shows a margin, such as plus or minus 5 for Gemini 4 Argon.

Is LMArena the same as Arena or arena.ai?

Yes. Arena was formerly LMArena, and the platform was previously known as Chatbot Arena, according to Arena's about page. Opening lmarena.ai on 15 November 2026 redirected to arena.ai. The leaderboards, voting and models are on the one site.

Is the LMArena leaderboard accurate?

It accurately measures what voters preferred, and it's not a test of correctness. Margins of 5 to 10 points mean the top places overlap, and newer models carry a Preliminary tag. Use it to shortlist models, then test two of them on your own prompt.

Is LMArena a benchmark?

Not in the usual sense. A benchmark scores models on a fixed set of questions. LMArena ranks them on live votes from people, and its FAQ says this makes the leaderboards grounded in community judgment, not static benchmarks. It measures preference, so pair it with a test-based score.

How often is the LMArena leaderboard updated?

Arena does not state a schedule on the pages we read, but each board shows its own date. On 15 November 2026 the text and WebDev boards were dated 11 November, text to image 12 November and vision 7 November. Check the date before you quote a rank.