OpenAI’s BioBasicsBench puts GPT-6 Astra at 66.7% on wet-lab tasks

OpenAI introduced BioBasicsBench on 6 November 2026, describing a 100-task benchmark developed with Ligo, a biotech company focused on enzyme design, to measure how well AI agents handle routine wet-lab analysis. At the xhigh setting, which allows more reasoning per task, GPT-6 Astra earned an average of 66.7% of the available credit, compared with 43.8% for GPT-5.6 Sol. [1]

What does BioBasicsBench test?

OpenAI published BioBasicsBench on 6 November 2026, a 100-task wet-lab agent benchmark built with Ligo. [1] OpenAI says existing benchmarks built around charts and images selected for scientific papers may miss the ambiguity of everyday lab work, including failed experiments and conflicting observations.

The 100 tasks cover five areas of routine wet-lab work: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. [1] Inputs include gel and blot images, plate photographs, microscopy stacks, sequencing reads, plate-reader exports and videos of bench work.

OpenAI says scientists created the tasks using data from their own experiments and paired each question with expert-written grading criteria. [1] Each task went through an internal review by a Ligo scientist who had not written it, an external review by a PhD-level wet-lab scientist, and a final internal review.

How did GPT-6 Astra score against GPT-5.6 Sol?

At the xhigh setting, GPT-6 Astra earned 66.7% of available credit, compared with 43.8% for GPT-5.6 Sol. [1] OpenAI says the results show progress on routine lab analysis, with room to improve on images, controls and context.

Agents return a structured answer with the requested conclusion, measurement or recommendation and, where required, supporting reasoning. [1] Answers receive credit against individual rubric criteria, so the benchmark can distinguish a correct conclusion from a partially correct analysis.

Why does OpenAI say everyday lab work matters?

OpenAI argues that misreading an image or overlooking a problem with experimental controls can lead scientists to the wrong conclusion and waste follow-up work. [1] In one illustrated example, GPT-6 Astra made colony-counting errors that included marking empty agar as a colony and leaving visible colonies unmarked.

OpenAI says tracking these capabilities will also help it understand implications for biosafety and biosecurity. [1] The post presents BioBasicsBench as a target for improvement rather than a claim that agents are ready to replace lab scientists.

Our read

Opinion: A dedicated wet-lab benchmark that scores raw gels and plates is more useful than another leaderboard built on cleaned paper figures, but a 66.7% headline still means these agents miss a large share of routine checks a working scientist should catch.

What we could not verify

• A visible publication-date line on the live OpenAI page in the text extract we obtained.
• A single human-expert baseline score for the full 100-task set.
• Scores for other frontier models that OpenAI did not highlight in the post.