LMArena AI review 2026: features, pricing, leaderboard and criticism

LMArena AI, now officially called Arena, is a free website where you type a prompt, two anonymous AI models answer, and you vote for the better reply. Those blind votes feed the best-known public leaderboard of AI models. It's genuinely useful as a free way to try frontier models side by side and as one signal of what people prefer, but it isn't a neutral IQ test, and the rankings have attracted serious, published criticism. I think that's the right way to read it.

Nobody at The AI Future ran hands-on tests for this review. The comparison rests on Arena's official pages and leaderboard data, the research papers that critique it, independent coverage and user discussions, as of late October 2026. All leaderboard positions below come from Arena's official leaderboard data published on 28 October 2026 unless stated otherwise.

About LMArena AI

It works like a blind taste test for chatbots. The project began as Chatbot Arena, a UC Berkeley research experiment launched on April 24, 2023, moved to the lmarena.ai domain in September 2024, incorporated as a company in April 2025 and rebranded to "Arena" on January 28, 2026, as Wikipedia records. Today lmarena.ai redirects to arena.ai, and the company is Arena Intelligence Inc. Its founders are Wei-Lin Chiang, Anastasios Angelopoulos and Ion Stoica.

So when people search "what is lm arena ai" or "ai arena", they mean the same thing, a public platform that does two jobs at once.

• It's a free multi-model chat app. You can chat with many proprietary and open models, compare them, build web apps and run agent tasks.
• It's also an evaluation business. Your blind votes become ratings on public leaderboards, and Arena sells deeper evaluations to AI labs and enterprises.

The company isn't a small research project any more. On October 8, 2026 Arena announced a $200 million Series B at a $3.1 billion valuation, co-led by Lightspeed Venture Partners and Khosla Ventures. In the same post it said it has passed $100 million in annualized revenue and counted 350 million sessions and 62 million votes across the platform.

The ranking method

Each vote is one game between two models. Arena's FAQ explains that it turns those games into ratings with the Bradley-Terry model, a statistical method for paired comparisons that is "similar to the Elo rating system" used in chess. It also adds extra features to the model for what it calls style control, which tries to stop longer or more heavily formatted answers from winning just for looking polished.

A few rules decide how far you can trust the numbers.

• Only anonymous votes count. Per the FAQ, votes cast after model names are revealed don't affect the leaderboard, and Side by Side or Direct mode chats (where you pick the models) aren't counted at all.
• Pre-release models appear under codenames. Labs test unreleased variants in Battle mode, and only a model that meets Arena's policy is added under its official name.
• Every score has an uncertainty band. Ratings come with confidence intervals, and new models with few votes have wide ones.

Worked example: what a 19-point gap means

On the style-controlled text leaderboard published 28 October 2026, Gemini 4 Carbon (High) scored about 1,528 and Claude Opus 5.5 (High) about 1,509, per Arena's official leaderboard dataset. On an Elo-style scale, the expected win rate for the higher model is 1 / (1 + 10^(−gap/400)). A 19-point gap gives 1 / (1 + 10^(−0.047)), or roughly 53%.

In plain terms: in a blind vote between those two, people would pick Gemini a little more than half the time. That's a real but small preference, and I wouldn't call it a different class of model. Gemini 4 Carbon also had only about 4,400 votes, against roughly 11,400 for Opus 5.5, so its confidence interval was wider (roughly ±9 points versus ±7).

Key features

Battle mode

It's the core feature. You submit a prompt, get two anonymous answers, and vote. Arena's how it works page notes that the model identities are revealed after you vote and you can keep chatting.

Side by Side and Direct chat

If you want a specific model, Side by Side lets you pick two to compare and Direct lets you chat with one. These are popular with people who want free access to frontier models, but the FAQ makes clear that the prompts are still collected for research.

Leaderboards for many tasks

Arena now runs separate leaderboards for text, WebDev and image-to-WebDev, vision, documents, search, text-to-image, image editing, text-to-video, image-to-video, video editing and agents. The text board alone breaks down into 29 categories in the official dataset, including coding, math, creative writing, multi-turn, many languages and industry slices such as legal or medicine.

Agent Mode and Agent Arena

Arena has pushed hard into agents in 2026. Its blog covers Agent Arena category and task-cost views (August 14), coding with GitHub in Agent Mode (August 24), a router called Max that it says matches frontier performance in Agent Arena with "37% cost savings" (October 12), and Connectors for Gmail, Google Drive, Linear, Supabase and other tools (October 27).

Factuality, AutoEval and alignment signals

Arena has started measuring more than preference. The same blog lists a factuality leaderboard (July 14), AutoEval scores that give provisional ratings before human votes pile up (July 30), and an Alignment Index launched with the Series B. Arena's own funding post admits why: it added factuality "after finding that the response people prefer isn't always the correct one."

Open data

Arena publishes its leaderboard data and several preference datasets on Hugging Face, which is where the rankings for this review come from. Its Series B post says it has open-sourced 375,000 data points.

Pricing and plans

For individual users, Arena is free. There's no public consumer pricing page (arena.ai/pricing returned a 404 on 31 October 2026), and the company makes its money from its commercial AI Evaluations service for model labs and enterprises, as its FAQ describes. A directory listing by AIToolsCoop also describes full free access to the public arenas and leaderboards.

Pricing checked October 2026.

The free plan costs you data. Conversations "may be used and shared," in the words of the FAQ, including in public research datasets and private evaluations with model providers, after de-identification. Arena itself recommends keeping sensitive information out of chats. The terms of use also require users to be at least 18. If you're wondering how a free product like this can still be a $100 million business, our explainer on how AI companies make money covers the pattern.

The leaderboard right now (published 28 October 2026)

The live leaderboard pages sat behind a bot check, so the same figures here come from Arena's official leaderboard dataset, last updated on 30 October with a leaderboard publish date of 28 October 2026. Arena's leaderboard changelog confirms Gemini 4 Carbon (High) was added to the Text, WebDev and Agent boards that day. Positions below follow score order; models with overlapping confidence intervals are probably tied.

Three things stand out in this table.

• Style control moves models. Meta's Muse Wonder (Max) sits 4th with style control and 8th without, so formatting seems to affect raw votes.
• Boards disagree. OpenAI's best text entry, GPT-6.1 Sol (Max), is only 27th on the style-controlled text board, yet OpenAI holds 2nd and 5th on WebDev. The top open-weight text model is Kimi K3 (Max) at 19th.
• Other boards have other leaders. In the same dataset, Claude Opus 5.5 (Max) leads the Agent board, Claude Fable 5 (High) leads Vision (published 19 October), and OpenAI's GPT-Image-2.5 Sunburst leads text-to-image (published 26 October).

Compare that with other evaluators. Artificial Analysis currently places Claude Opus 5.5 first across all six of its industry capability indexes, while one of Scale's leaderboards puts Gemini-4-Argon on top. Different tests, different winners. If you want a refresher, see our guides to what an AI model is and how LLMs work.

The criticism

Arena-style rankings have been criticised in detail, and I'd read the main arguments before quoting a leaderboard position.

Private testing and selective disclosure

The most cited critique is "The Leaderboard Illusion" (April 2025), by researchers from Cohere, Stanford, Princeton, MIT, Ai2 and others. It argues that "undisclosed private testing practices benefit a handful of providers" who can test many variants and publish only the best, and identified 27 private variants tested by Meta before Llama 4. It also estimated that Google and OpenAI received 19.2% and 20.4% of all arena data, while 83 open-weight models combined received 29.7%. Cohere's Sara Hooker told BetaKit that the benchmark "has shown a clear pattern of unreliability in rankings."

To be fair to Arena, the company disputed the paper, saying its pre-release testing policy had been public and listing what it called factual errors, BetaKit reports. A LayerLens summary of the exchange notes Arena's position that every developer may submit multiple variants. Separately, after Meta submitted a chat-tuned Llama 4 Maverick that differed from the public release, Arena updated its policies.

Style over substance

The second critique is that casual voters reward confident, well-formatted answers. Surge AI, a data-labelling company that sells its own evaluation services, published a blunt post, "LMArena is a cancer on AI", saying it reviewed 500 votes and "disagreed with 52% of them." The commercial angle matters, but Arena's own move to style control and factuality scores concedes part of the point, I think.

Vote rigging

A 2025 paper, "Improving Your Model Ranking on Chatbot Arena by Vote Rigging", simulated attacks on 1.7 million historical votes and found rankings could be moved "by rigging only hundreds of new votes." The authors also tested defences and called for continued work on them.

Conflict of interest

Arena now earns revenue from evaluation work for labs whose models appear on its leaderboard. Startup Fortune argues this creates a conflict "the company can't wave away," while noting Arena says submissions are processed blindly and its voter pool is too large to manipulate easily.

The defence of Arena is also worth hearing. One widely shared rebuttal calls it "the single best single benchmark available today" for broad human preference, while warning that you should never rely on any one leaderboard to pick a production model.

User and reviewer reactions

About 35 comments were read across four Reddit threads from January to May 2026 that discuss the leaderboard. The split was roughly a quarter positive, a third mixed and under half negative. Supporters value the blind human-preference signal; one commenter in a thread comparing eval sites wrote "I trust lmarena more" than composite benchmarks. Critics argue it measures taste rather than capability and is less relevant to agent work; in another thread a user wrote that its "days as the bellwether are long past." On Trustpilot, the unclaimed lmarena.ai profile has just two reviews, too few to mean much.

Best suited to

If you mainly want free model access, our list of the best free AI models covers other routes. If you've already narrowed it down, our Claude AI review and ChatGPT with GPT-6 review go deeper on the two most common finalists.

LMArena AI alternatives

• Artificial Analysis: independent intelligence, speed and cost comparisons based on its own evaluations. Better when price per task matters.
• OpenRouter rankings: ranks models by tokens processed through its API by developers. It shows what people actually pay to use, not what they prefer in a blind test.
• LiveBench: a contamination-focused benchmark from Abacus AI that grades long-horizon work such as reports, bug fixes and apps. Its 2.0 release landed on 25 October 2026.

Verdict

LMArena AI is probably the best free place to compare frontier models blind, and its leaderboards give a fast, large-sample read on what people prefer. I'd treat it as one input and no more. A small gap such as 19 points means a 53/47 preference split, style and formatting still move votes, and the private-testing and conflict-of-interest critiques haven't gone away. Check the confidence intervals, look at the category that matches your work, and run your own prompts before you commit. Skip it if you need privacy for your chats or a direct measure of accuracy.

FAQ

What is LM Arena AI?

LM Arena AI (now Arena, at arena.ai) is a free platform where you compare two anonymous AI models on your own prompt and vote for the better answer. The votes produce public leaderboards for text, coding, vision, images, video, search and agents.

What is the LMArena AI platform used for?

People use it to try many AI models for free, compare them side by side, and check human-preference rankings. AI labs also use it to test unreleased models under codenames and to buy deeper evaluations.

Is LMArena AI free?

Yes, for individual users. Arena earns money from its AI Evaluations service for labs and enterprises. The trade-off is that your prompts and votes can be shared, after de-identification, in research datasets and with model providers.

Which model is number one on LMArena?

On the text leaderboard published 28 October 2026, Gemini 4 Carbon (High) was first, with Gemini 4 Argon (High) second and Claude Opus 5.5 (High) third. On WebDev, Claude Opus 5.5 (Max) was first. Rankings change often, so check the live board.

Is the LMArena leaderboard reliable?

It's a reliable measure of what its voters prefer, with published confidence intervals, but not a direct measure of accuracy or capability. Researchers have raised concerns about private testing, sampling imbalance, style bias and vote rigging, and Arena has responded with policy changes, style control and factuality scores.

Is LMArena the same as Chatbot Arena?

Yes. Chatbot Arena became LMArena in 2024 and was renamed Arena in January 2026. The old lmarena.ai address now redirects to arena.ai.