OpenAI introduced BioBasicsBench on November 6, 2026, to test AI analysis of lab evidence. Developed with biotech company Ligo, it uses 100 tasks from real experiments, OpenAI said.
The company tests data analysis, not hands-on lab work, and reports both progress and errors. Its announcement identifies no peer-reviewed paper.
What BioBasicsBench scores measure
OpenAI awards partial credit for individual parts of an answer. Thus, the scores below do not show the share of experiments completed correctly. Each model setting had five attempts per task.
BioBasicsBench: selected models at xhigh reasoning effort
Source: OpenAI; date shown is publication date
Data as of: 2026-11-06
OpenAI reports the highest score for GPT-6 Astra among these three models. Scores measure partial credit, not fully solved tasks.
| Model | Available credit earned |
|---|---|
| GPT-6 Astra | 66.7% |
| GPT-6.1 Sol | 62% |
| GPT-5.6 Sol | 43.8% |
The test covers seven OpenAI models. However, its results remain the developer’s own assessment; the announcement provides no independent replication.
Reading a gel is more than spotting a pattern
For example, BioBasicsBench uses gel images, according to OpenAI. The National Human Genome Research Institute explains the underlying lab method, called electrophoresis. An electric current moves molecules through a gel. This sorts them by size and charge.
The gel acts as a sieve: smaller molecules travel faster than larger ones, the institute explains. Scientists run samples of known sizes on the same gel as an unknown sample to gauge its size. In other words, the reference sample gives the image its scale.
OpenAI’s examples show why context matters. Astra noticed a control plate used the wrong antibiotic, but in another task it chose the wrong colonies despite seeing their patterns correctly.
What a lab would still need to check
The National Institute of Standards and Technology (NIST) says accuracy tests should reflect the conditions in which an AI system will be used. Its AI Risk Management Framework also sets out what reliability means. A system must work as required, over time, in the conditions set for it.
NIST also says checks should account for errors that cause greater harm. All errors do not carry the same risk. It also says human help may be needed when a system cannot spot or fix its own mistakes. That guidance offers a way to assess a lab assistant beyond its average score.
For deployed systems, NIST describes ongoing testing or monitoring to check that they still work as intended. OpenAI says BioBasicsBench will help track gains in lab analysis.
