An Arena runs one prompt on several harness and model pairs side by side.

What it is

An arena sends one prompt to several harness and model pairs at once and shows them side by side, so you can compare the work, the time it took and what it cost before you choose a pairing for your product.

Each column is an ordinary task, with its own session, files and cost. The arena only records which tasks belong together.

Running one

• Open Arena under Build in the console.
• Add up to ten columns, each a harness and a model, and send the prompt. Every column starts at once.
• Message every column at once to keep going, or stop them all.
• Each column shows its time and cost once its turn finishes. Pick a winner, or open any column as its own task and keep working in it.

The live demo

The demo at app.harnessrouter.ai/arena/demo replays two real runs of five columns each, an SVG drawing and a contract review, with their recorded times, costs and output files played back in about twelve seconds. It starts no tasks and costs nothing. Sign in to open it.

What it costs

The arena itself is free. Each column is billed as its own task, so an arena of five columns costs what those five tasks cost. During the Free plan's 7-day trial, before you add prepaid credit, two tasks run at a time, so a third column is refused until one finishes. Prepaid credit or a paid plan removes the limit.

From the API

These routes save the grouping and start nothing. Run each column as a task through POST /v1/responses and record its session id in the column. An id that does not exist answers 404 arena_not_found.

Deleting an arena removes the grouping; its tasks stay where they are. A deleted arena is removed 30 days later.

POST   /v1/arenas         {"name": "Contract review", "columns": [{"id": "col_1", "harness_id": "...", "model": "...", "session_id": "..."}]}
GET    /v1/arenas         the arenas in the workspace
GET    /v1/arenas/{id}
PUT    /v1/arenas/{id}    replace it, with every column it should keep
DELETE /v1/arenas/{id}