Last updated: 2026-10-04
Built on more than 10 million user evaluations, according to TechCrunch, Arena (formerly LMArena) is a free web platform where people send one prompt to two anonymous AI models and vote for the better answer. Those votes feed public leaderboards for text, image, video, coding and agents.
About Arena (formerly LMArena)
Arena began on 24 April 2023 as Chatbot Arena, a research project at UC Berkeley, and became LMArena when it moved to its own domain in September 2024. Its co-founders are Anastasios Angelopoulos (CEO), Wei-Lin Chiang (CTO) and Ion Stoica, the Berkeley professor who advised the project before it incorporated in April 2025. It closed a $150 million Series A on 6 January 2026 and dropped the "LM" on 28 January 2026, so the service now trades as Arena at arena.ai. The old lmarena.ai address serves the same site.
The core loop is simple. In Battle Mode you send one prompt to two anonymous models, vote for the better reply, and only then see which models answered. Side by Side lets you pick the two models yourself, Direct chats with one, and Agent Mode runs models through longer tasks that use tools. The leaderboard tabs cover agents, coding agents, WebDev, agents for work, text, image and video. Labs such as OpenAI, Google, Anthropic and DeepSeek supply models. Recent launches that surfaced there include Claude Fable 5.1, Gemini 4 Argon, GPT-6 Sol and GPT-6 Astra, while Claude Opus 5 also sits on the agent board.
The public site is not where the money comes from. Revenue comes from AI Evaluations, launched on 16 September 2025, which sells evaluation work to model labs and enterprises. TechCrunch reported a $100 million annualized run rate in June 2026 and noted that Arena bills for consumption, so the figure is not recurring revenue in the usual sense. The company says its leaderboard will stay free as a public service and that neutrality, breadth and open datasets are standing commitments.
Read a ranking as a measure of what voters preferred. In April 2025 a version of Llama 4 Maverick that differed from the public release placed well, and the platform updated its policies afterward. If you are choosing a chatbot for a job rather than for bragging rights, the best AI chatbots guide works from tasks. Calling many models from one API is what OpenRouter sells. Browse its neighbours in the conversational and voice AI group when you want to compare answers before you commit.
Screenshots
Pricing
Checked against the Arena terms of use (updated 23 February 2026) and the AI Evaluations launch post on 4 October 2026. The public site is free of charge, and the terms keep the right to charge for features later. AI Evaluations, the commercial service for labs and enterprises, has no price published in its launch post; TechCrunch describes billing as consumption-based.
| Tier | Monthly price | What it includes |
|---|---|---|
| Public Arena | Free | Battle Mode, Side by Side, Direct chat, Agent Mode, public leaderboards; some features need an account |
| AI Evaluations | Consumption-based, no published price | Evaluation work for labs and enterprises using community feedback, auditable feedback samples, service-level agreements on delivery |
Key Features
- Battle Mode: Two anonymous models answer the same prompt, you vote for the better reply, and the identities are revealed only after the vote.
- Side by Side and Direct chat: Side by Side compares two models you choose yourself, while Direct chats with a single model at a time.
- Seven leaderboard views: Tabs rank agents, coding agents, WebDev models, agents for work, text, image and video, so a model is judged on the job you care about.
- Agent Mode with task cost: Models run longer sessions with tools such as web reading and a shell, and a cost view plots score against median cost per task over the last 14 days.
- AutoEval scores: Introduced on 30 July 2026, AutoEval gives calibrated ratings on real tasks before enough human votes have piled up.
- Pre-release model testing: Labs test unreleased models on the platform under codenames, which is how early versions of GPT-5 and Nano Banana appeared before launch.
Pros
- The public leaderboard is free, and the company commits in writing that it will stay a public service.
- Blind voting on real prompts answers questions a fixed test set cannot, such as which reply reads better to a person.
- Unreleased models show up here before launch, so it can show early reception when vendors publish nothing.
Cons
- The terms let Arena share your inputs with the model providers, who need not keep them confidential, so private or client material does not belong here.
- A vote records a preference, not correctness on your task, and a study reported by Fast Company found that hundreds of rigged votes could shift rankings.
- Terms confine you to personal or internal business use, with reselling of the service or its output barred.
- The service is web-based, and an App Store search for LMArena on 4 October 2026 returned apps from unrelated developers rather than the vendor.
Data Handling
- Training-data policy
- The terms of use (23 February 2026) let the company and AI providers use your content to improve their services, and inputs are shared with providers who need not keep them confidential.
Frequently Asked Questions
What does Arena cost in 2026?
Nothing for ordinary use, so there is no plan list to compare. The paying customers are labs and enterprises that order AI Evaluations. Its launch post publishes no rate, and TechCrunch says billing follows consumption, which means the bill tracks how much evaluation work a customer orders.
Is LMArena (now Arena) free to use?
Yes. The terms say Arena currently offers the service free of charge and reserve the right to charge later. Registration is optional, but some features need an account, and you must be 18 or old enough to form a binding contract. The terms also confine use to personal or in-house business needs.
What are Arena's closest competitors?
TechCrunch says Arena has no direct rival on the public side, and Yupp, a similar crowdsourced startup, shut down in March 2026. For trying many models yourself, [Poe](/hub/tools/poe) offers chat and [OpenRouter](/hub/tools/openrouter) offers an API. For open models you can download and run, see [Hugging Face](/hub/tools/hugging-face). On the paid side, the company says it competes for the same budget as human-labeling firms Mercor, Surge and Scale AI.
What separates Arena from OpenRouter?
Arena is where you compare answers and vote, and it publishes the resulting rankings. [OpenRouter](/hub/tools/openrouter) is where you send paid requests to many models through one API. Choose Arena to learn which model people prefer, and OpenRouter once you need to call the model from your own code.
What does it take to start using Arena?
Open arena.ai in a browser, type a prompt in Battle Mode and vote on the two anonymous answers. The leaderboard pages can be read without signing in. Skip private, client or regulated material, because the terms let Arena pass your inputs to the model providers.
Top Alternatives
- OpenRouter: Pick Arena to see which models people prefer for free; pick OpenRouter when you need to call those models from your own code.
- Poe: Pick Arena to compare answers blind and read public rankings; pick Poe for a chat app you use daily.
- Hugging Face: Pick Arena to learn how people rate a model's replies; pick Hugging Face to download, host or fine-tune open models.