Market · AI in practice

BioBasicsBench measures how AI reads laboratory evidence

The OpenAI and Ligo evaluation combines routine tasks and raw data, with results showing progress alongside remaining errors.

Company or topicBioBasicsBench

What the sources report

On November 6, 2026, OpenAI introduced BioBasicsBench, developed with Ligo. The evaluation contains 100 laboratory tasks and uses evidence such as microscopy images, colony plates, and instrument readings.

According to OpenAI, GPT-6 Astra earned 66.7% of the available credit at the highest reasoning-effort setting cited in the table. GPT-5.6 Sol earned 43.8%. These figures describe performance on the evaluated tasks.

The company itself emphasizes that higher scores do not yet demonstrate reliable judgment in the laboratory. Examples of failure include reversed colony identities and counting errors. Interpreting an image requires attention to what the evidence actually supports.

My take

I see value in a concrete evaluation for tracking this progress. It helps locate mistakes and guide improvements, while preserving researchers’ review of the evidence when making decisions about the results.

Author

Adriano Meirinho · Co-founder & CAIO (Chief AI Officer) at Celcoin. About

All notes