> ## Content Index
> Fetch the complete content index at: https://soundbytes.news/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenAI test asks AI agents to read lab gels, plates and microscope images
- URL: https://soundbytes.news/openai-test-asks-ai-agents-to-read-lab-gels-plates-and-microscope-images/
- Published: 2026-11-06T20:22:28.000Z
- Updated: 2026-11-18T08:49:43.000Z
- Description: OpenAI says BioBasicsBench, built with the biotech company Ligo, tests whether AI agents can read raw lab evidence and spot errors. The scores come from OpenAI's own tests of its models.
- Author: Echo
- Tags: Tested and measured

One AI model noticed that a control plate in a lab test carried the wrong antibiotic label: kanamycin, not the ampicillin used. Another model saw no colonies on the same plate and based its conclusions on that. A control is a check on whether an experiment's results can be trusted.

The two models are GPT-6 Astra, which rejected the plate as a reliable control, and GPT-5.6 Sol. Both are OpenAI's, and the example comes from [OpenAI's post](https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science?ref=soundbytes.news) introducing BioBasicsBench, dated 6 November 2026\. A benchmark is a fixed set of tasks used to score AI systems. OpenAI says it developed this one with Ligo, a biotech company focused on enzyme design.

It has 100 routine tasks from wet labs, places where scientists run experiments with biological samples. The test is for AI agents, which are AI systems set to work on a task. The inputs are raw evidence: gel and blot images, plate photographs, microscopy stacks, sequencing reads and videos of bench work. One simple task is reading the molecular weight of a DNA sample in a gel.

OpenAI says existing tests use charts and images chosen for scientific papers, which may miss the ambiguity of real lab work. It says scientists wrote the tasks from their own experiments. Each was reviewed by a Ligo scientist who had not written it, an outside PhD-level wet-lab scientist and a final internal review. Answers earn credit against an expert-written rubric.

OpenAI tested seven of its own models. A setting is one combination of a model and a reasoning level, and OpenAI ran 37 of them, with five attempts on each of the 100 tasks, for 18,500 results. At the "xhigh" setting, which allows more reasoning per task, GPT-6 Astra earned an average of 66.7% of the available credit, against 43.8% for GPT-5.6 Sol.

OpenAI gives other examples. In microscope images where cells came in and out of focus, Sol sought a biological explanation, while Astra recognised a z-stack, a series of images taken at different depths. Sol also labelled cells as dead from a dye signal that could include light from a second dye. Astra flagged the ambiguity and called for better controls.

OpenAI also shows where Astra still goes wrong. Counting bacterial colonies on a plate, it merged close neighbours, missed visible colonies and marked empty agar, the jelly the bacteria grow on. That gave an overcount of about 12% against the benchmark reference.

In a DNA test of colonies, Astra picked the wrong colonies. That would lead a scientist to purify the wrong plasmid, a small loop of DNA used in cloning. OpenAI writes: "Higher benchmark scores do not yet mean models can make reliable judgments in the lab."

The post says progress on these tasks could help scientists spend less time troubleshooting. OpenAI adds that tracking these abilities will help it understand implications for biosafety and biosecurity.

## Sources

- [Measuring how ai agents handle wet lab science](https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science?ref=soundbytes.news) · openai.com