LMArena

advancedwidespread

A public website where users anonymously choose between responses from paired AI models, generating a continuously updated ranking based on the votes.

In plain terms

LMArena is like a blind test at a wine competition: two glasses are placed before you without labels, you taste both, and you say which one you liked better. Here, the two glasses are the responses of two language models to the same question, and only after your vote is it revealed which company's model you saw.

A ranking list is compiled from these millions of votes, which news outlets often cite when writing about 'which AI is best right now'. Developer companies also monitor it, as a high ranking carries significant advertising value.

Traditional tests measure performance on pre-written task sets with a scoring key. LMArena instead measures which answer real humans prefer in response to real questions. This is more realistic but also more subjective: it measures preference, not necessarily correctness.

How it works

You enter a question or task, the system selects two models, and sends it to both.
You see the two answers side by side without the models' names, so the brand does not influence you.
You vote: left is better, right is better, a tie, or both are bad. The model names appear only after this.
Based on the results of the matches, the system calculates a ranking using a statistical method similar to the Elo rating known from chess: whoever defeats a strong opponent receives more points.
Separate leaderboards are created for different areas, such as text, programming, image understanding, image generation, or web search.

An everyday example

You read in an article that a new model "took the lead on the LMArena leaderboard." This means that in blind tests, people chose its answers more frequently than those of competitors in the recent period. You can try it yourself: ask your own question and vote.

Who is affected

everyoneFor free, without brand names, you can compare which model answers your own questions better.

entrepreneur, executiveIt is a good starting point for model selection, but does not replace testing on your own tasks.

developerProgramming and other domain leaderboards help narrow down which models are worth testing for an application or a coding agent. → Code

marketerRanking-based advertising claims can only be correctly interpreted if you know what the leaderboard measures and what it does not. → Marketing

A common misconception

MythThe top-ranked model on LMArena is the best for everything.

In realityThe ranking shows which answer was generally preferred by voters. For your specific task, such as Hungarian legal text or accounting, a lower-ranked model may perform better.

Limits and risks

Voters often value the appealing, confident, well-formatted answer more than the accurate one because they rarely verify the facts.
Small differences between scores are often within the statistical margin of error, so a few places' difference at the top does not mean much.
In 2025, researchers criticized that large companies can test multiple model versions non-publicly and publish only the best one, which may distort the rankings. Platform operators disputed several of these claims.
Voters do not represent the average user: most ask in English and are tech-interested, so you learn little about Hungarian-language performance.

Similar, but not the same

benchmark (performance test) – Traditional benchmarks measure on fixed tasks with pre-defined correct answers, while LMArena ranks based on human subjective votes. In a broader sense, LMArena is also commonly called a benchmark.
Live Score – The Elo score is the calculation method itself, while LMArena is the platform that applies this type of scoring to AI models.
chatbot – The chatbot is a single service you converse with, while LMArena is a testing platform for comparing multiple models.

Related concepts

Broader

AI model evaluation, Benchmark, Ranglista

Narrower

Vakteszt, pairwise comparison, Live Score

On the edge

Large language model, chatbot, Human evaluation, Open-source model, hallucination

A brief history

The platform launched in 2023 as Chatbot Arena, an open project by researchers at the University of California, Berkeley (LMSYS group). In 2025, it became an independent company, operating under the name LMArena. In early 2026, it shortened its name to Arena, but the name LMArena continues to live on in common usage.

Where things stand

By the end of 2026, one of the most cited AI rankings, regularly referenced by developer companies and the press, is increasingly known simply as Arena. Meanwhile, due to criticisms of its methodology, more people are interpreting it alongside other metrics.

Status: 2026 November

In the news

Appears in 2 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Agencies and vendors launch AI visibility tools and servicesAug 10, 2026 · Marketing
Fake AI job candidates pass interviews and vanish; Deel buys deepfake-defense startup ClarityAug 3, 2026 · Work & HR

Also known as

LMArena · LM Arena · Chatbot Arena · LMSYS Chatbot Arena · Arena · lmarena.ai · LMArena-ranglista · LMArena leaderboard

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.