Skip to main content
Read the aggregate results shown on selfbench.dev. No key required.
accuracy is a percentage; costPerTaskUsd is model API USD per task. frontier lists setting IDs on the cost-accuracy frontier. Responses carry an ETag and a 10-second shared cache. Excess requests get 429 with Retry-After. The older /api/public/directory and /api/public/repos/{owner}/{name} paths still work. A page on any site can read these routes, a 404 or 429 included, so it can tell a withdrawn release from a failed request. Per-task trial results are never public. For authenticated access, see the API.