A quiet colosseum for student work.

Study Arena is a blind head-to-head for AI models, scored on the work students actually do: planning an assignment the night it gets set, checking a source, fixing a limp thesis, working out what you need on the final, drafting the message you have been avoiding, making a page look like someone designed it.

You pick a category and type one prompt. The server quietly picks two models from the enabled pool and runs them at the same time through the identical harness — the same system prompt, the same four tools (fetch a page, run a calculation, check the date, render a webpage), the same eight-iteration tool cap, the same output budget. Their answers stream side by side as Contestant A and Contestant B, with no names attached.

When both lanes finish, you vote: A, B, tie, or both bad. Only then do the identities appear, along with what your vote did to each model’s Elo. Every vote updates two tables — the category you battled in and the overall standings — which is how a model can be the best study explainer on the board and middling at research.

Why the identical harness matters

Most model comparisons are really comparisons of scaffolding: one model gets better tools, a longer prompt, more retries. Here the harness is a single shared module and both lanes call it with the same arguments. What you are voting on is the model, not the wrapper around it.

Prior art

The blind-battle-plus-Elo format was popularised by Chatbot Arena (LMSYS), and Design Arena applied it to generated interfaces. Student Arena is an independent project inspired by that lineage, built for student tasks and for models working with tools. It is not affiliated with either.

Fine print

• No accounts, no rate limiting beyond a light per-IP throttle, one vote per battle.
• Chat conversations live in your browser tab only — nothing is persisted.
• Battles and ratings are stored in a local database on whichever machine is running this app.
• Model-generated webpages render in a sandboxed iframe with no same-origin access.