# OpenAI's new BioBasicsBench tests whether AI can read messy wet-lab data. GPT-6 Astra scores 66.7%

> The 100 tasks use real gels, plates, microscopy and sequencing data from scientists' own experiments. Astra beats GPT-5.6 Sol, but still makes mistakes that would sink an experiment.

- URL: https://devisdead.ai/news/openai-s-new-biobasicsbench-tests-whether-ai-can-read-messy-wet-lab-data-gpt-6-a
- Published: 2026-11-08T23:31:13Z
- Author: DevIsDead Newsroom
- Section: AI
- Tags: openai, biology, benchmarks, science
- Originally reported by Unite.AI (https://www.unite.ai/openai-introduces-biobasicsbench-to-test-ai-agents-on-wet-lab-data/).

Lab work isn't the clean charts in published papers. It's blurry images, failed controls and conflicting results. OpenAI's new BioBasicsBench, built with enzyme-design company Ligo, tests whether AI agents can handle that mess, according to [Unite.AI](https://www.unite.ai/openai-introduces-biobasicsbench-to-test-ai-agents-on-wet-lab-data/), whose article was written by an AI research agent and reviewed by its editors.

## The benchmark

Its 100 tasks were written by scientists from their own experiments, across protein methods, imaging and cytometry, assays, nucleic acid methods and cloning. The evidence is the raw stuff of a lab bench: photos of gels, blots and colony plates, stacks of microscope images, sequencing output, plate-reader files, and even videos of protocols being run. Rubrics reward correct numbers, sound reasoning, and recognizing when the evidence isn't enough to conclude anything. Each task went through three reviews, including an outside PhD wet-lab scientist.

At the highest reasoning setting, GPT-6 Astra scored 66.7% and GPT-5.6 Sol 43.8%, with Astra using about half as many output tokens per task.

## Where Astra improved

OpenAI's examples show it catching context Sol missed: noticing that two dyes' signals could overlap and asking for controls; rejecting a negative-control plate labeled for the wrong antibiotic; and recognizing that cells "disappearing" were just a z-stack captured at different depths, not a biological event.

## Where it still fails

OpenAI is clear that higher scores aren't reliable lab judgment. On one PCR task, Astra picked the wrong lanes by assuming a bigger product meant success, a mistake that would lead to purifying the wrong plasmid. It overcounted colonies by about 12%, and sometimes noted the right observation, then ignored it in its final answer. OpenAI says tracking these abilities also feeds its biosafety assessments.

## Why it matters

This is a better kind of benchmark: real data, honest rubrics that reward "I can't tell," and published failures. The gap it exposes, seeing the evidence but not following it to the conclusion, is the same one showing up in [agents doing open research](/news/nine-ai-agents-tried-real-follow-up-research-for-national-lab-scientists-none-cl). Helpful lab assistant, yes. Unsupervised lab tech, not yet.

**[Scoreboard](/scoreboard) update:** Draw. AI gets much better at reading lab results, and still picks the wrong plasmid.

Stay undead.