What is LLM Arena
Arena began as a research project at the University of California, Berkeley, and was called Chatbot Arena. Then the platform became LMArena, and on 28 January 2026 the team dropped “LM” from the name: now it is Arena. In search people still call it LLM Arena or the AI model arena.
The idea is simple: models do not solve a fixed set of known tasks but answer questions from real people. The questions are new each time, and the judge is a person who does not know which model they are looking at.
According to Arena itself, as of January 2026 more than 5 million people a month from 150 countries use the platform. Model developers and annual reports such as the Stanford AI Index cite its tables.
How one comparison works
You write any question; there are no ready tasks. The platform picks two models and shows their answers side by side, without names.
Then you vote: which answer is stronger, a tie, or both bad. If the winner is not clear from the first answer, you can continue the chat and vote later.
Model names appear only after the vote. So that a model does not give itself away, the authors filter out, by keywords, chats where the answer mentions the name of a model or a company.
Vote collection began in April 2023. By January 2024 there were about 240,000 votes from more than 90,000 people, with comparisons among more than 50 models.
How votes become a ranking
Each vote is the outcome of one match between two models. From thousands of matches you need one number for each model, like a chess player’s rating.
At first Arena used the Elo system: after each match, the winner gains points and the loser loses them. In December 2023, with more than 130,000 votes, the platform moved to the Bradley-Terry model.
The Bradley-Terry model is a statistical way to estimate each participant’s strength from the whole match history at once, not game by game. It assumes that a model does not change over time, and it gives stable scores with a confidence interval, the range where the true rating most likely lies.
Since August 2024 Arena has separated a model’s strength from the style of its answer. Style features were added to the calculation: answer length and the number of headings, lists and bold text. The authors write that style clearly affects the places, and its main factor is length.
What Arena rankings exist
Arena has not one table but several boards by task type. As of March 2026, the platform’s menu lists text, website building (WebDev), vision, documents, search, agents, images from a description, image editing, video from text and from an image, and video editing.
To choose a language model, people usually look at the Text board: general questions and conversation. If you need a model for website code, look at WebDev; for working with files, Document.
Places on different boards do not match: a model can rank high in text and low in video or code. So the phrase “first in Arena” says nothing without the name of the board.
Why you should not trust the ranking blindly
In May 2025 the paper “The Leaderboard Illusion” came out. The authors studied about 2 million Arena comparisons: 243 models from 42 developers from January 2024 to April 2025.
The first claim: large developers test private variants of a model on the platform before release and show the one with the top score. One developer tested 27 variants before a model came out. If you take the top result of many tries, the place ends up higher than the model’s real strength.
The second claim: data is shared unequally. By the authors’ estimate, Google got about 19.2% of all Arena data, OpenAI about 20.4%, and 83 open models together about 29.7%. The authors show that this access to data leads to fitting the quirks of Arena, not to better quality overall.
The third correction comes from Arena itself: style, above all answer length, affects the vote. A vote shows what a random visitor liked, not what is right for your task.
How a company should choose a model: Arena plus its own test
Arena works as a first filter: it cuts weak models and shows which ones are worth testing. After that, a test on your own tasks decides.
Collect a few dozen real questions from clients or employees with correct answers. Run them through 3–5 models from the top of the board and compare the answers blind, as Arena does: without model names. Remember the lesson about style: judge the facts, not the length and formatting.
Besides quality, a company cares about the price per request, speed, the Russian language, work with your documents and where the data is stored. How to build such a test is covered on the page on agent evaluation, and the term “benchmark” in the glossary explains why a general task set does not replace your own test.
How Arena differs from task-based tests and composite indexes is covered separately: how AI model rankings work and which ones to trust.
Where to see the current places
Places in Arena change with every new model, so we do not repeat the table here. A snapshot of the places, with the date of the check and a link to the board, is on the LLM ranking page.
For general context, the Stanford AI Index 2026 is useful. According to it, in the 2026 Arena ranking the top US model scored 1503 points and the top Chinese model 1464. The gap between the top rows is small, and the choice between them comes down to price, access and your own test.
If your question is not “who is first” but “what to use for the task”, start with the page on which AI to choose for text, code and documents. Other materials on language models are collected in the LLM section.