OpenAI launches BioBasicsBench: 100 real wet-lab tasks to stress-test AI agents
OpenAI · x · 2026-11-07
OpenAI, together with enzyme-design biotech Ligo, released BioBasicsBench, a benchmark measuring how well AI agents handle the routine analyses of wet-lab science.
- It draws on 100 tasks from real experiments, testing whether agents can interpret raw lab data, spot errors, and recognize when evidence doesn't support a conclusion
- The accompanying chart shows GPT-6-class models (Astra, Sol) clearly ahead of the GPT-5.x family, with GPT-6 Astra reaching nearly 67% mean reward around a 6k-token budget, while GPT-5.x models cluster between 25%–45%
- OpenAI frames the benchmark as a way to track how model gains translate into scientific progress, biosafety, and biosecurity
Related event: OpenAI Releases BioBasicsBench for Wet-Lab AI Agents(3 posts)→
More from Models
- Inkling 2S matches Dots3 on ARC but costs roughly 3.5x more — teortaxesTex · 2026-11-07
- Open ASR Leaderboard adds private Egypt and Nigeria conversational data for accent breakdown — andimarafioti · 2026-11-07
- Users say Claude Opus 5.5 remains unmatched a month after release — prasenx · 2026-11-07
- Ripple-65B Preview: Apache 2.0 open US model tops TerminalBench 2.1 with $32M R&D — The Decoder · 2026-11-07
- Mirage: A 22MB VLM Runs Detection, OCR and Segmentation in One Pass at 74ms on Phone NPU — Henrie_the_dreamer · 2026-11-07
- 4B Nemotron 3 Nano claimed to beat frontier models on task-specific workflows — eliano · 2026-11-07