Trying to pick the best AI chatbot in 2026 is hard. Benchmarks disagree, vendors cherry-pick results, and every release claims a new record. One project cuts through that noise: lmarena ai, the site most people still call Chatbot Arena.
It runs head-to-head votes between two anonymous models and turns the results into a public ranking.
This guide is chatbot arena explained from the ground up: what lmarena ai is, how its leaderboard is built, how to vote, and where it falls short.
The LLM News treats lmarena ai as the closest thing the field has to a live, crowd-sourced scoreboard. The project began in May 2023 as Chatbot Arena, a research tool from LMSYS and UC Berkeley SkyLab. It has since grown into a product with its own company behind it.
Today the lmarena ai site draws millions of users each month and records tens of millions of conversations. Everything on the public board comes from real people voting on real answers, which makes it the closest thing to a primary source in model evaluation. Here is how it works.
Before the deep dive, a note on scope. This guide covers the live board, the voting flow, the fairness debate, and how a science benchmark called LLMSRBench compares with a chat arena. If you are picking tools for your own work, start with our LLM tools guide, then come back here for the lmarena ai rankings.
- Launched: 3 May 2023, by LMSYS and UC Berkeley SkyLab.
- Now: Arena at arena.ai, run by Arena Intelligence Inc.
- Reach: 5 million-plus monthly users and about 60 million conversations per month in January 2026.
- Text board: 379 model configurations and roughly 8 million text votes in early October 2026.
- Cost: free to use, no account needed.
What Is lmarena ai and What Was Chatbot Arena?
Chatbot Arena launched on 3 May 2023. It came from LMSYS, a multi-university nonprofit, together with UC Berkeley SkyLab.
The setup was simple: show a user two anonymous answers to the same prompt, ask which one is better, then collect enough votes to build a ranking with no vendor marketing mixed in.
The researchers wanted a preference test that ordinary people could run, so they built one that runs on everyday users instead of lab staff.
For years people called it the LMSYS arena. LMSYS is a 501(c)(3) nonprofit with roots at Berkeley, Stanford, UC San Diego, CMU, and MBZUAI. Its projects include Vicuna, FastChat, SGLang, MT-Bench, and the LMSYS-Chat-1M dataset.
Chatbot Arena is marked “graduated” on that list, which means it spun out into its own product while the rest of the group keeps building inference and evaluation tools. If you want the background first, read what an LLM actually is.
The name changed twice. In September 2024 the site moved to the lmarena.ai domain and people started saying LMArena. In April 2025 the team incorporated as Arena Intelligence Inc.
On 28 January 2026 it rebranded again as Arena at arena.ai. Many visitors still search for lmarena ai and land on imarena.ai, an alternate spelling of the same product.
So the question “imarena ai vs lmarena” has one answer: one product, one leaderboard, two addresses. Nothing about your votes or the ranking changes depending on which domain you type.
The product also answers a practical question. Lab benchmarks show how a model scores on fixed questions, and vendors can tune for those questions until the number looks good.
A blind head-to-head vote on your own prompt is harder to game, which is why product teams watch the lmarena ai board closely whenever they ship a release.
How the lmarena ai Leaderboard Is Ranked
Every vote is pairwise. Two models answer the same prompt and one wins. Those results feed a Bradley-Terry model, the same math used to rank tennis players from match results, and it outputs an Elo-style rating for each entry.
A higher rating means the model wins more often against the field. You do not need the formula to read the board, but you should know that ratings depend on who beat whom, not on raw scores from a fixed test.
Three rules shape the lmarena ai board. Only votes cast while both models were anonymous count, because identities reveal after the vote. Every entry carries a 95% confidence interval, so a two-point gap is noise rather than a lead.
The board also tracks model configurations instead of brand names: in early October 2026 the lmarena ai text board listed 379 model configs and about 8 million text votes, and that snapshot may change.
Scale explains the trust. The site passed 5 million monthly users and roughly 60 million conversations per month in January 2026. In 2026 the first models ever crossed 1500 Elo, a threshold nobody reached in the three years before it. For a plain read of the current order, open the live text leaderboard and sort it yourself.
Style Control is the toggle most people miss. A raw ranking rewards long, heavily formatted answers, because voters like clean markdown and generous structure.
Style Control runs a regression that discounts answer length and markdown formatting, then reorders the list. The moves are big: claude-opus-4-1-thinking climbed from rank 3 to rank 1, while glm-4.5 fell from rank 2 to rank 12.
Read the lmarena ai list without the toggle and you are reading a style contest alongside a substance contest.
Reading the board takes two minutes. Find your model, check the interval beside it, and compare the overlap with the entry above. If the bars overlap, the two are effectively tied.
Toggle Style Control once to see how much formatting is doing for each answer. The lmarena ai leaderboard makes more sense when you treat every position as a range instead of a fixed rank.
The top of the table moved fast in 2026. As of early October 2026, a snapshot that may change, Gemini 4 Argon sat at about 1525, Claude Opus 5.5 at about 1507, Claude Opus 4.6 at about 1504, and Claude Fable 5 at about 1504.
The strongest open-weight entry, Kimi K3, reached about 1488. For contrast, the September 2025 archive shows gemini-2.5-pro at 1469 with 4,048,204 votes across 243 models. New releases move these numbers weekly, so we fold the shifts into today’s AI news roundup.
How to Vote on Chatbot Arena in Four Steps
Voting is free and takes under a minute. You do not need an account, and there is no cap on how many times you can try. The lmarena ai voting page is plain on purpose: one prompt box, two answer panes, four buttons.
- Go to arena.ai and type any prompt you like.
- Two anonymous answers appear side by side, labeled only as left and right.
- Vote for the left answer, the right answer, a tie, or both bad.
- Identities reveal after your vote. Only votes cast before that reveal count toward the rating.
That last rule matters. Seeing which model produced an answer changes how you judge it, so the arena discards those votes. Good prompts are specific and checkable: ask for a short code fix, a fair comparison, or a fact you already know. Vague prompts like “write something nice” produce two similar answers and teach the lmarena ai ranking very little.
Votes also stack up quickly. Tens of millions of conversations run through the site each month, and only the blind portion feeds the rating, so the board keeps absorbing fresh preferences as models update.
When a vendor ships a point release, the change shows up in the order within days instead of months.
Is Chatbot Arena Reliable? The Main Criticisms
No public ranking is neutral, and the arena has real critics. The sharpest paper aimed at it is The Leaderboard Illusion (arXiv:2504.20879), presented at NeurIPS 2025. It reports that Meta tested 27 private variants before releasing one.
It also finds that 64 to 66 percent of silently deprecated models are open-weight, and that the top two providers received around 40 percent of arena data. Teams with that data access saw their ArenaHard win rate rise from 23.5 percent to 49.9 percent.
The other complaints are structural. Answers get judged on style as well as correctness. Votes are binary, so a narrow win and a blowout look identical.
The people who vote self-select, and the heaviest users are the ones who enjoy the game, streaks included. Critics also note that a single lmarena ai score hides how uneven a model’s performances can be from one topic to the next.
Treat the lmarena ai ranking as one signal among several. Pair it with objective tests, read the confidence intervals, and check model releases in our programming news section when a version is quietly swapped out.
Automated relatives of the arena help too: MT-Bench scores 80 multi-turn questions with GPT-4 as judge on a 0 to 10 scale, Arena-Hard-Auto separates models 87.4 percent of the time and agrees with arena outcomes 89.1 percent of the time for about $25 per run, and AlpacaEval covers single-turn replies. That mix answers “is chatbot arena reliable” with evidence instead of a yes or no.
None of that makes the data useless. The arena still aggregates far more real conversations than any lab test, and the anonymous phase blocks simple stuffing. The honest reading: lmarena ai tells you what people enjoy talking to, while the papers tell you what happens behind that conversation.
LLMSRBench vs Chatbot Arena: Two Different Benchmarks
LLMSRBench is not a chat arena. Its name expands to LLM-SRBench, where SR stands for symbolic regression. The work comes from arXiv:2504.10415 and was accepted as an ICML 2025 Oral.
It holds 239 problems drawn from Chemistry, Biology, Physics, and Materials, and it is built to defeat memorization of famous equations. A model cannot win by recalling a formula it saw during training; it has to derive the right relation from the data it is given.
That makes llmsrbench vs chatbot arena a category question. LLMSRBench measures objective scientific accuracy against ground truth. The lmarena ai arena measures what people prefer in a conversation.
One grading is right or wrong, the other is a matter of taste, and a model can top one table while sitting in the middle of the other.
| Chatbot Arena (lmarena ai) | LLMSRBench | |
|---|---|---|
| What it measures | Which of two answers people prefer | Whether a model derives correct equations |
| Score | Elo-style rating from Bradley-Terry votes | Accuracy across 239 science problems |
| Source of truth | Subjective human votes, blind and anonymous | Objective ground truth in four sciences |
| Track record | Running since May 2023 at arena.ai | ICML 2025 Oral, arXiv:2504.10415 |
| Best for | Choosing a daily assistant | Testing scientific and symbolic reasoning |
Use them together. Ask the arena which assistant people enjoy, then ask a science benchmark whether that assistant is actually right. For hands-on practice and new model versions, our LLM updates feed tracks every release as it lands.
Pick by job. If you want a daily writing or coding partner, start with the arena and the style control toggle. If you need a model for equations, forecasting, or lab work, put LLMSRBench and similar science checks first. Both kinds of number belong in the same decision.
The short version: lmarena ai is the best window we have into how models behave for real people, and it is still a preference poll. Watch the style control toggle, read the intervals, keep the science benchmarks in view, and check back as new models enter the field.
Frequently Asked Questions
What is lmarena ai?
lmarena ai is the product most people still call Chatbot Arena. Two anonymous chatbots answer your prompt, you vote on the better one, and those votes build an Elo-style leaderboard. The site now lives at arena.ai and stays free to use, with no account required.
How does chatbot arena work?
Your prompt goes to two models picked at random, and neither one is labeled. You vote for the answer you prefer, the labels appear, and only the votes cast before that reveal are fed into the Bradley-Terry ranking that drives the board.
Is chatbot arena reliable?
It is a strong crowd signal, not a final verdict. Critics point to style bias, binary votes, self-selected voters, and unequal data access between providers. Read the confidence intervals, use Style Control, and pair the arena with objective benchmarks before you decide.
Is imarena ai a different product from lmarena ai?
No. imarena.ai is an alternate spelling domain of the same Arena product, not a second leaderboard. Votes, rankings, and model entries are shared across the addresses, so imarena ai vs lmarena comes down to which one you typed.
Who is on top of the lmarena ai leaderboard?
In the early October 2026 snapshot, which may change, Gemini 4 Argon led at about 1525 Elo, ahead of Claude Opus 5.5 at about 1507. The strongest open-weight entry, Kimi K3, sat at about 1488 on the text board.
What is llmsrbench compared with chatbot arena?
LLMSRBench tests symbolic regression on 239 science problems and scores objective accuracy in chemistry, biology, physics, and materials. Chatbot Arena scores subjective human preference between two chat answers, so the two benchmarks never compete for the same title.
