Two frontier models answer your prompt blind — click the better one, no card, no plan
Two frontier models answer your prompt blind — click the better one, no card, no plan▓▒░ two answers, names hidden ░▒▓ ▸ 
your click decides
The models people pay $20 a month each for sit on one page here — two at a time, names hidden — and the whole price of admission is clicking which answer was better.
Start here — 15 seconds, nothing to install, nothing to pay
- Open https://arena.ai — the older address https://lmarena.ai still works, same arena
- Hit Battle, paste the prompt that’s been failing you
- Two answers come back with no names on them. Click the better one.
That click is the rent. Nothing else gets asked of you.
Why it’s free — and why that isn’t a catch
Not a leak, not a key, not a cracked build. It is a research rig run by LMSYS — the place that used to be called LMSYS Chatbot Arena.
The trade is plain once you see it:
Sponsored compute is the supply. OpenAI, Google, Anthropic, Meta, Alibaba and the rest hand over API access, because an impartial rating from here carries more weight than their own numbers.
You get the models. They get the votes. Nobody is selling you a subscription.
What that gets you that a $20 plan doesn’t
- Your prompt, not a benchmark. Paste your own stack trace, your own SQL, your own chart screenshot — and make the frontier models answer that.
- Names hidden until you vote. You judge the answer instead of the logo on it.
- Free access, no plan to buy. Same models as the paid tiers, nothing to subscribe to.
- Experimental builds sit in the same list. The models labs are still testing show up beside the public ones — the pickers below are the proof.
Five arenas, not one chat box
 arena |
what you throw at it |
|---|---|
| Text & reasoning | the everyday prompts — GPT, Claude, Gemini, DeepSeek class models |
| Coding & web | your bug, your refactor, code that renders a running interface |
| Vision & multimodal | upload an image — chart reading, diagrams, design analysis |
| Image & video | text-to-image and text-to-video prompts, compared side by side |
| Hard prompts & math | the academic logic and STEM problems that break normal chat |
Battle or direct chat? Battle hands you two anonymous answers — that is where you find out what is genuinely better. Direct chat lets you name the model when you already know which one you want.
Who is actually in the picker
Ten model lists, straight off the site. Read the version tags before anything else.
the names off every pricing page
the labs with no plan to sell you
Half of those version numbers are not for sale anywhere yet. That is what a blind arena is for.
🧯 What not to paste — and the habits that make it pay
Do
- Use it for complex debugging and code review — tricky error logs, edge cases, stack traces, two diagnoses side by side
- Test refactoring and performance work — how each model rewrites for speed, readability, memory efficiency, clean architecture
- Lean on the Code and Vision arenas — autocomplete logic, unit test generation, architecture and diagram analysis
- Benchmark for your own stack — Python, C++, Rust, React, SQL — then keep the model that actually wins
Don’t
- Paste API keys, credentials or secrets. None of them, ever
- Drop .env files, tokens or database connection strings into a prompt
- Upload proprietary or enterprise code, client projects, or anything under NDA — keep non-public work out
- Trust the output blind — review, lint, test and sanitize before anything reaches production
- Waste time guessing in battle mode when you need one specific model — switch to direct chat
The models are free because your vote is the product — spend it on your own problem.
!