Skip to content

Benchmark suite

The Benchmark pane runs a registry of categorized tasks — reasoning, code, refactoring, and more — against the models you pick, scores the results, and lets you export them. It’s the “measure it again next month” counterpart to the Model Council.

  1. Open it with open benchmark suite.
  2. Pick a category (or a whole suite) from the registry.
  3. Pick the model(s) — installed agent CLIs work out of the box; OpenRouter models join once a key is set.
  4. Run and watch per-task results fill in.

The two-model compare view lines both up on the same tasks so differences are visible per item, not just in the total.

Results export to a file for records or sharing — useful for before/after checks when a CLI updates its default model.

The artifact categories go beyond SVG and HTML: Artifacts · Music (Web Audio) asks models to ship a self-contained page that synthesizes a melody or drum loop in code (no audio files), and Artifacts · Motion (canvas/CSS) asks for animated scenes — a bouncing canvas ball, a CSS-keyframe storyboard. Answers render live in the sandboxed preview (click the play button in the preview to hear music — browsers require a click before audio), and the auto-check chip reads the capability at a glance: Web Audio music, Animated, or Audio + motion. No external generation APIs are involved — the model’s own code is the instrument.

  • Benchmarks consume real usage/quota on the agents they run — a large suite on a metered CLI counts against it.
  • The registry is categorized so you can benchmark only what you care about (e.g. just code tasks).