Chatbot Arena
Arena Intelligence Inc. · Other · Chatbot
Chatbot Arena, now called Arena, shows two anonymous AI models answering the same prompt and lets your vote set a public leaderboard. It's free and covers text, image, video and coding.
About Chatbot Arena
What Is Chatbot Arena
Chatbot Arena is a website where you talk to two AI models at once without knowing who they are. You type a prompt, both models answer, and you pick the response you prefer. Only after you vote do you see which company made each answer. Every vote feeds a public ranking that tracks how models compare.
The project started in 2023 as a research tool from a team of UC Berkeley researchers, who wanted a better signal for AI model comparison than fixed exam-style tests. Static tests can be gamed, leaked or outgrown, while real people asking real questions are harder to prepare for, and that honesty is what makes the LLM leaderboard worth reading. That idea turned into a company, Arena Intelligence Inc., which now runs the site at arena.ai and uses the old lmarena.ai address as a redirect.
The setup comes with tradeoffs you should know. Ratings reflect what a crowd of mostly anonymous users preferred, not what an expert panel measured, and the crowd skews toward people who enjoy browsing new AI tools. Votes are also easy to cast casually, so a single vote carries little weight. Arena earns money by selling evaluations to model labs and enterprises, which is a conflict critics raise, and the company answers by publishing its method and shared datasets.
Getting Started
- Open arena.ai in a browser. You can send a prompt right away, and signing in is only needed to keep your history and unlock scheduled evaluations.
- Choose a mode. Battle serves two anonymous models and hides their names. Side by Side lets you pick both models yourself. Direct Chat opens a single model for a normal conversation.
- Write your prompt, and use the tool icon to switch between text, image, video or a build task such as a landing page.
- Read both answers, then vote for the better one. Names appear the moment you submit, and only anonymous votes count toward rankings.
- Check the leaderboard afterwards to see how your pick places on the text, vision, search, image, video or WebDev board.
Product Information
A quick look at Chatbot Arena's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- People choosing an AI subscription
- Writers and developers comparing output styles
- AI followers tracking the field
- Product teams picking a model to build on
Tasks
- Testing a prompt against several models
- Checking whether a new release lives up to its launch post
- Getting a quick read on image or video models
- Comparing coding output for a small build
Scenarios
- A family debate about which assistant is smartest
- Choosing between two paid plans
- A developer pitch meeting
Key features
Anonymous Battle Mode
Battle is the core of the site. You send one prompt and get two answers with no model names attached. Voting reveals both identities at once, which makes it harder to favor a familiar brand. Only votes cast while the names are hidden count toward the leaderboard, so the ranking stays blind.
Bradley-Terry Rankings
Scores come from a Bradley-Terry model, a statistical method built for paired comparisons and close in spirit to chess ratings. Instead of an exam score, a model earns standing by beating others in head-to-head votes. The site also applies style control, which factors out preferences for longer or more polished answers so length alone doesn't win.
Category Leaderboards
There isn't one single table. Text, WebDev, Vision, Search, Text to Image, Image Edit, Text to Video, Image to Video, Video Edit and Agent each get their own board. That matters because a model that tops creative writing can fall behind on code, and picking by overall rank alone hides that gap.
Pre-Release Codenames
Model labs hand over unreleased versions for blind testing, and they often appear under a codename before launch. You get hands-on time with something months before it ships, and labs get feedback while they can still change things. If a model passes the public release rules, its score later appears on the board under its real name.
Side by Side and Direct Chat
Not every task needs a vote. Side by Side lets you name both fighters when you already know what you want to compare, and Direct Chat opens one model for plain use. Both collect prompts for research, but neither changes the leaderboard, which keeps the blind board clean.
Build, Image and Video Tasks
The site goes past text. You can ask a model to build a landing page, a dashboard, a small game or a storefront, generate an image or a short video, or edit what it produced. Each output family has its own leaderboard, so a model that writes well but draws badly won't hide behind one score.
Open Data and Published Methods
Arena releases a large share of its voting data on Hugging Face and publishes papers on its method, from the original Chatbot Arena research to prompt-level rankings. You don't have to take the ranking on faith: you can read the data, the style control work and the leaderboard policy behind it.
Pros and cons
Pros
- The blind vote removes brand bias, so a smaller lab's model can reach the top on the strength of its answers.
- Free access to frontier models, including pre-release versions under codenames, is hard to match anywhere else.
- Separate boards for text, vision, search, image and video mean you can check the specific skill you care about.
- Published data and method papers let researchers and buyers verify the ranking rather than trust a marketing page.
- Every vote takes seconds, so a rough comparison is only a few minutes away.
Cons
- Rankings reflect crowd preference, not a careful test of correctness, so a confident wrong answer can score well.
- Votes are unordered and easy to rush, and researchers have published critiques of gaming and prompt leakage, so treat scores as a signal, not a verdict.
- Arena earns revenue from model labs, which means the company that ranks models also sells them evaluation services.
- Prompts you type are collected and some are released in research datasets, so anything sensitive belongs elsewhere.
Frequently asked questions
Chatbot Arena is a free website where two anonymous AI models answer the same prompt and your vote decides which one responded better. Those votes build the public leaderboard that ranks models across text, images, video, search and coding.
Related content
Explore related tools, skills, and articles for Chatbot Arena.
Chatbot Arena Alternatives
Odyssey 3
Odyssey · OtherOdyssey-3 is Odyssey's foundation world model. It simulates physics in real time, turns a prompt into an interactive environment, and adapts to robots, cars and games.
JarvisCore
Prescott Data · Coding · Productivity · OtherJarvisCore is an open source Python runtime where AI agents work as equal peers on an agent mesh, claim tasks from a shared ledger, and reach external services through a zero-trust broker.
Clef
Cloudflare · Business · OtherClef is Cloudflare's open-source decision model for Workers AI. It returns calibrated probabilities for typed questions, not generated text, so agents can classify, route, and act fast.