OpenAI Unveils BioBasicsBench for Wet-Lab AI Agents
The 100-task benchmark, built with biotechnology company Ligo, tests how AI agents interpret raw laboratory evidence.
Updated on Nov. 8, 2026 in Life Sciences
OpenAI introduced BioBasicsBench on Nov. 4, a benchmark of 100 tasks designed to assess how AI agents interpret raw wet-lab experimental material. Ligo, a biotechnology company, built the tasks from its own experiments and prepared the grading rubrics.
Why it matters
Most AI evaluations rely on charts and images selected for scientific publication, which may not reflect the messy data scientists handle at the bench. A benchmark built from raw gel images, colony plates, microscopy data and instrument readouts offers a more realistic test of whether AI agents can support laboratory work.
The evaluation covered seven models across 37 model and reasoning-effort configurations, with each configuration attempting every task five times for 18,500 total results. At the xhigh setting, GPT-6 Astra led with a 66.7% mean reward, while GPT-5.5 recorded 35.4%.
The players
OpenAI
OpenAI is an artificial intelligence research company that develops large language models, including the GPT series evaluated in this benchmark.
Ligo
Ligo is a biotechnology company that built the benchmark's tasks from its own experiments and prepared the grading rubrics.
The details
BioBasicsBench draws its exercises from gel images, colony plates, microscopy data and instrument readouts rather than curated publication figures. Each task was reviewed by a Ligo scientist, an external PhD-level wet-lab scientist and an internal reviewer before being included.
Timeline
- Nov. 4: OpenAI introduces BioBasicsBench.
The Big Picture
BioBasicsBench shifts AI evaluation from curated scientific figures toward raw experimental material, a step beyond general-purpose GPT benchmark scores. The results show a wide spread between models, with GPT-6 Astra scoring 66.7% against GPT-5.5's 35.4% at the xhigh setting.
Researchers evaluating AI tools for laboratory work can use benchmarks like this to judge whether an agent can handle raw gel images, plates and instrument readouts rather than only polished figures. Readers following AI-assisted research should expect model choice and reasoning settings to noticeably change results.
The takeaway
A benchmark built from raw experimental evidence gives a clearer picture of AI readiness for lab work than curated publication figures do. Scores vary sharply by model and reasoning setting, so specific configurations should be checked before relying on them.
Further reading
For related coverage, explore the Life Sciences section.
Source note: This article includes information reported by TokenPost.
