ResearchOpenAI

OpenAI launches BioBasicsBench to test biology AI

Original title: Measuring how AI agents handle everyday wet-lab science

Plain language

What it is: OpenAI introduced BioBasicsBench, a benchmark with 100 tasks to measure how well AI agents interpret raw experimental data from everyday biology lab work. The benchmark was developed in collaboration with the biotech company Ligo.
How it works: It tests AI agents on routine laboratory tasks using real experimental data, such as images, genetic sequences, and videos. The agents must analyze this data to identify errors, interpret results, and recognize when the evidence is insufficient to draw a conclusion.
Where it is available: Not stated in the source.
Example: An AI agent can be asked to read the molecular weight of a DNA sample from a gel image, or notice if a laboratory plate is labeled with the wrong antibiotic.

What it is: OpenAI introduced BioBasicsBench, a benchmark with 100 tasks designed to evaluate how AI agents interpret raw experimental data and perform routine wet-lab science tasks. It was developed in partnership with Ligo, an enzyme design biotech company.
How it works: The benchmark feeds multimodal inputs like FASTQ sequencing files, flow cytometry files, microscopy z-stacks, and protocol videos to AI models. Models must output structured answers containing conclusions, measurements, or recommendations with supporting reasoning, which are then graded against expert-written rubrics.
Where it is available: Not stated in the source.
Example: A model is evaluated on its ability to recognize that a series of microscope images showing cells appearing and disappearing is a z-stack depth scan, instead of looking for a biological explanation.

The headline and explanation on this page are based on OpenAI's post. The full details are in the original.

Measuring how AI agents handle everyday wet-lab science