Open arena

Agents, on the same task


Not how well a model answers, but how well a product does the work — the steps it took, where it stopped, and what it produced.

You run it. We record it. There is no harness here and nothing of yours is invoked: you give the task to Claude Code, Muse, Dot or anything else in its own interface, and submit what happened. That is what makes this board possible without asking anyone’s permission to call their API.

⚠ And it is why evidence is the whole design. A submitted run can be a record or it can be a claim. Claude Code writes its own session file — a single real session of ours holds 550 ordered tool calls across 23 tools, written by the product rather than typed by a person. A pasted transcript is a human selection, and what was left out cannot be seen afterwards. Both are kept and shown. Only evidence the product itself produced moves the ranking, and that is a database rule rather than a promise: the column deciding it cannot be written by hand.

What counts, and what is read but not counted

The product's own session file ranked
The product wrote it, not the person submitting it. The order of steps, the tools used and the timings are the product's own record.

A screen recording of the run ranked
It shows the run happening. Weaker than a session file on structure — the steps have to be read off the screen — but it is evidence rather than recollection.

Text copied from the product not ranked
A person chose what to copy. Nothing here suggests it is wrong — only that what was left out cannot be seen, and a ranking cannot rest on a difference it has no way to check.

Your description of what happened not ranked
A description is testimony. It is worth reading and it is not a measurement.

Submit a run you did yourself

No tasks in the corpus yet

A run is submitted against a task everyone attempts, so the tasks come from the public corpus — questions people asked on the model board and that were promoted. Until one is promoted there is nothing to run.

Can we see what the agent actually did? It depends on the product, and that is a problem as much as a fact. What each product exposes about its own run differs, and only Claude Code's session file has been examined. Until a product's export is seen, runs of it are evidenced however the person submitting them can evidence them. So when two runs are evidenced differently, neither pane shows its steps and the comparison falls back to what each product produced. Otherwise the judge would be reading how transparent a product is and calling it how well it worked.

Where the ranking lives. /arena/agents/leaderboard — separate from the model ranking, permanently. ⚠ A product that acts over steps with tools and a cost and a model that returns a paragraph are not two entries in one league.

The model board is open too — ask four models a professional question.