OpenAI releases BioBasicsBench benchmark; GPT‑6 Astra scores 66.7% on wet‑lab tasks
OpenAI and biotech firm Ligo introduced BioBasicsBench, a new benchmark of 100 routine wet‑lab tasks that require AI agents to read raw experimental data such as gel images, colony plates, microscopy stacks and instrument readouts. The benchmark evaluates whether agents can interpret the evidence, spot errors and recognize when the data do not support a conclusion.
Key points
- BioBasicsBench benchmark contains 100 wet‑lab tasks across five experimental areas.
- GPT‑6 Astra achieved 66.7% mean reward at xhigh setting, versus 43.8% for GPT‑5.6 Sol.
- Evaluation covered seven models, 37 settings, 18,500 attempts, showing multimodal reasoning improvements.
Seven OpenAI models were tested across 37 reasoning‑effort settings, with five attempts per task for a total of 18,500 results. At the highest "xhigh" setting, GPT‑6 Astra achieved an average of 66.7% of the available credit, compared with 43.8% for GPT‑5.6 Sol, showing notable gains in multimodal reasoning, error detection and context awareness. The results highlight both progress and remaining gaps, such as occasional over‑interpretation and mis‑counting of colonies.
The authors argue that tracking performance on BioBasicsBench will help improve AI assistance for biological research, reduce time spent on troubleshooting, and inform biosafety considerations.
Model pages: GPT-5.6 Sol → · GPT-6.1 Sol → · GPT-6 Luna →
Measuring how AI agents handle everyday wet-lab science
OpenAI · 6 November 2026
Loading the full article…
This text was published by OpenAI. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- PSI proves Aaronson-Ambainis conjecture using Anthropic’s non-public model · 1 src
- Mit develops CheMeleon model that outperforms other methods · 1 src
- Baseten releases Serverless Loops for open-model fine-tuning with per-token billing · 1 src
- Insilico Medicine releases open-source Longevity Grok AI system with 30 agents · 1 src
- Ro launches research lab to analyze patient data for clinical insights · 1 src
Comments
via GitHub Discussions