Atlas

Benchmarks

← All benchmarks

Robo-bench — Lego Manual (public)

Agents · 2026-10-13

Public Robo-bench arm task: build a five- or six-brick model on a studded baseplate from an illustrated manual, pressing bricks home. Scores 15 per brick placed as shown and 100 for the finished model. A single public, single-seed episode nested in the arm suite, so it is a diagnostic.

Top models (higher is better)

ModelScore
Claude Opus 5.515.0
GPT-6.1 Sol15.0
GPT-6 Astra15.0
Claude Fable 5.10.0
Gemini 4 Argon0.0
Grok 4.70.0
Muse Wonder Full0.0
Loading Atlas data…