The News, Explained
On November 6, 2026, OpenAI introduced BioBasicsBench, an AI evaluation it built with biotech company Ligo. It tests how well agents read raw evidence from labs that work with biological samples, including microscope images and instrument readings. The 100 tasks ask models to spot errors and recognize ambiguous data from real experiments, alongside interpreting evidence that is clear enough to support an answer. Source
Both models attempted the same 100 tasks five times each at xhigh, a setting that allows more reasoning per task. OpenAI reports an average of 66.7% for GPT‑6 Astra and 43.8% for GPT‑5.6 Sol. Graders award credit criterion by criterion for things like the conclusion, a measurement, or the reasoning behind it. So 66.7% is the average share of available credit, including partial credit. It cannot be converted into 67 fully correct answers out of 100. Source
One task asked the models to judge whether cells were dead based on a fluorescent signal. GPT‑5.6 Sol classified the cells as dead from their brightness. Astra noticed that the detector could pick up light from another dye and called for a comparison that could separate the two signals. Astra still made basic mistakes elsewhere. On one image of a plate, it struggled to distinguish nearby colonies, or clusters of growing cells, and marked some empty areas. It counted about 12% more colonies than the benchmark reference on that image. That figure describes a single example. Source
A separate OpenAI study in 2025 tested model-proposed changes through repeated physical experiments. BioBasicsBench tests how models read evidence that already exists. The two studies measure different kinds of work, so their numbers do not compare directly. Source Source
OYOPICK’s Take
A promising first job for an assistant is to flag missing evidence before anyone settles on a conclusion. Spotting the need for a better comparison, as in the fluorescence example, could help a researcher decide what to test next. But as the colony example shows, a wrong count can throw off any decision built on it. If an answer shows which observations it used and what it could not distinguish, a researcher may be able to find the part worth checking before repeating the experiment. Source