BenchmarksRun an evaluation

Run an evaluation

Pick a test, choose an AI Runtime model, and score a small sample. No local setup and no API key in the browser.

Checking hosted runner configuration…

ONE CLICK, BOUNDED

What happens when you run

  1. Pinned tasks

    Tasks come from the immutable dataset revision recorded on the benchmark page.

  2. Server-side model

    AI Runtime is called with a server-side credential. Your browser never receives the key.

  3. Exact, unpublished score

    You get the correct count, sample denominator, task outcomes and provider-reported cost.

Need the full benchmark?

A prefix sample is a quick signal, not a leaderboard result. Export every recipe, configuration and missing setting for a full local run.

Build a full run plan →

The hosted runner supports MMLU-Pro, AIME 2025 and AIME 2026. More adapters are added only after their schema and scorer fixtures are verified.