Arena

They are one product under three names. The team announced on 28 January 2026 that LMArena is now Arena at arena.ai, dropping the LM. The About page describes a platform created by UC Berkeley researchers and previously known as Chatbot Arena, and the company behind it is Arena Intelligence Inc. Older articles and links may use either earlier name.

There is no price list. The address arena.ai/pricing returned a page saying it could not be found when checked, and no plan table appears on the site. The only statement about cost is in the Terms of Use: the company currently offers the service free of charge and retains the right to charge for the service or any features of it.

So free is a present condition, not a promise. Arena says it sustains itself by selling AI evaluations to enterprises, model labs and developers, and it publishes no price for those. Treat any figure from a third-party directory as unconfirmed, and re-read the terms before relying on free access for regular work.

Arena gives each user a daily balance of usage credits that can be used for free. The first prompt starts a 24-hour usage period, an indicator appears in the text bar as the limit approaches, and after it is reached a message shows a countdown to the refresh. Text and Code Arena Battle Mode usage does not count against the balance.

The size of the balance is not published, so it cannot be converted into a number of prompts. The account help page says logged-in use has higher rate limits than logged-out use, again without figures. Expect limits to change and check the in-app indicator instead of planning around a fixed allowance.

Battle Mode is the default on the homepage. Direct, Side by Side, Image Arena and Video Arena require login, according to the account help page. The modes differ in what happens to your vote, which matters if you care about contributing to the rankings.

• Battle Mode: two anonymous models; names appear after the vote; only votes cast while names are hidden count toward official rankings.
• Side by Side: two models you choose; matchups are not anonymous and votes do not count toward the public leaderboard, though prompts are collected for research.
• Direct: one model at a time with no voting; prompts are still collected for research.
• Agent Mode: a multi-step workflow at arena.ai/agent that can use web search, image generation and a sandbox, with file uploads and a downloadable workspace.

Arena leaderboards rank models by how often they are preferred in head-to-head votes. The FAQ says ratings come from the Bradley-Terry system, similar to Elo, with extra adjustments such as style control. The help page lists columns for rank, rank spread, a score with a confidence interval, vote count and token price. Tabs cover text, code, vision, document, image, search and video, with expert, occupational and factuality filters.

Ranks, scores and model positions are left out here because they change often and depend on the filter chosen. The Agent leaderboard is built differently, from single-agent session signals and causal tracing rather than pairwise votes. Rankings measure preference among voters on the prompts they typed, so check a model on your own task before choosing it.

Under the terms you consent to Arena sharing your content with third-party AI providers to complete requests, and those providers may not be required to keep it confidential. You keep ownership, but you grant Arena a broad, perpetual license to use it, including for improving the service and for marketing. The FAQ adds that conversations can appear in public research datasets and in private evaluations with model providers, after de-identification.

Arena advises against including sensitive information, and the privacy policy says the service is not intended for anyone under 18. To remove a chat, archive it, then delete it from the archive; the help page calls this irreversible and says removal from systems typically completes within 30 days, with limited data kept where law or security requires.

Agent Mode accepts PNG, WebP, JPEG, PDF, GIF, plain text, Markdown, CSV, HTML, XML, CSS, JavaScript and JSON uploads, while Battle Mode accepts PNG, JPEG, WebP and PDF. Without a connected repository, files sit in a session workspace and download as a zip. With a GitHub repository, the assistant commits to a working branch and opens a pull request.

Once that pull request is merged or closed, the session can no longer push, and later files cannot be downloaded or carried into a new session, so push work first. Uploading an image in Text Arena switches you to Image Arena for editing.

1. Note the credit indicator before the first prompt.
2. Send the prompt in Text Arena Battle Mode, vote, and note the indicator again.
3. Repeat the same prompt in Direct mode with one named model.
4. Note when the indicator appears and whether a refresh countdown is shown.

Accept the credit help page description if Battle Mode leaves the balance unchanged and Direct mode reduces it. Reject or note any mismatch with a date. No credit size or limit is reported until the steps are run.

To compare assistants you would use outside Arena, see ChatGPT and Claude. For cited web answers as the main job, see Perplexity.