OpenAI and Ligo introduce BioBasicsBench: 100 wet-lab data-analysis tasks. At xhigh, GPT-6 Astra earns 66.7% mean reward versus GPT-5.6 Sol’s 43.8%, OpenAI reports; scores allow partial credit.

OpenAI’s original BioBasicsBench chart of mean reward by model and reasoning effort versus output tokens, with confidence intervals.

Published 6 Nov 2026 · 19:38 UTC

OpenAI — BioBasicsBench ↗ · Source dated 6 Nov 2026

Browse all Live updates · Read Features