OpenAI BioBasicsBench tests AI agents on 100 wet-lab tasks; GPT-6 Astra scores 66.7%

What OpenAI released

OpenAI published BioBasicsBench on 6 November 2026, a benchmark that asks whether AI agents can interpret the raw evidence produced by ordinary laboratory work. It was built with Ligo, an enzyme-design company, and has 100 tasks in five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning.

The inputs are what a bench scientist actually sees: gel images, colony plates, microscopy, cytometry data and instrument output files. According to coverage of the post by R&D World and Unite.AI, each task was reviewed by a Ligo scientist, an external PhD-level wet-lab scientist and an internal OpenAI reviewer. The benchmark covers analysis only. Pipetting, running instruments and robotic execution are outside its scope.

Answers are scored with partial credit, and results are reported as the average share of available points. OpenAI ran seven of its own models in 37 combinations of model and reasoning effort, with five attempts per task in each setting, about 18,500 scored attempts in total.

The task types are the daily checkpoints of a molecular biology lab. Reading a gel means judging whether a band sits at the expected size and whether the controls behaved. Counting colonies on a plate tells a scientist whether a transformation worked. Colony PCR picks which clones carry the right insert, and a wrong pick sends days of downstream work after the wrong sample. Errors at these steps are cheap to make and expensive to discover late, which is why the benchmark rewards careful reading of the controls as much as the main result.

How the models scored

At the xhigh reasoning setting, GPT-6 Astra scored 66.7%, with a 95% confidence interval of 60.3% to 73.1%. GPT-6.1 Sol followed at 62.0%, then GPT-5.6 Sol at 43.8%, GPT-6 Luna at 38.3%, GPT-5.5 at 35.4%, GPT-5.4 at 33.2% and GPT-5.2 at 31.2%. The top score is a little more than twice that of GPT-5.2, the oldest model tested, so the gap between generations on this kind of work is large even inside one company's lineup.

OpenAI's own caveat is the more useful number. The company says higher scores do not yet amount to reliable laboratory judgment, and the post walks through failures. Coverage highlights a colony count that was about 12% too high, and a colony PCR task where the best model chose the wrong samples to carry forward. Several write-ups note a recurring shape: the model describes the image correctly and then draws the wrong conclusion from the controls, which in practice could mean purifying the wrong DNA.

These are OpenAI's results on OpenAI's models. The post reports no human baseline and no scores for models from other labs, and outside groups have not yet reproduced the numbers.

Why a lab-data benchmark matters

Most science benchmarks test knowledge or reasoning about text, such as exam questions or paper summaries. Wet-lab work is messier: failed experiments, ambiguous bands on a gel, contamination and conflicting readouts. A benchmark built on raw images and instrument files is closer to the task labs would actually hand an agent, which is triage of daily results before a scientist decides the next step.

It also sits next to OpenAI's biosecurity work. OpenAI rates its frontier models under its Preparedness Framework, and GPT-6 Astra was its first model rated Critical for cybersecurity. BioBasicsBench measures routine analysis rather than dangerous capability, but the same skill, reading experimental evidence well, matters for both scientific use and risk assessment. OpenAI published a separate wet-lab study in December 2025.

The release follows other lab claims this autumn, including an OpenAI model-written preprint on the Exponential Time Hypothesis. Benchmarks that ship with failure cases are easier to judge than headline claims.

What to watch next

The open questions are availability and comparison. The coverage reviewed here does not show a public download of the tasks, rubrics or scoring code, so other labs cannot yet run their models on the same set. A human expert baseline would show how far 66.7% is from a trained scientist. Watch for a dataset release, results from Google DeepMind or Anthropic models, and whether OpenAI adds the benchmark to future system cards. More coverage is in the AI research section.

Questions people ask

What is BioBasicsBench?

A 100-task benchmark from OpenAI and Ligo that tests whether AI agents can interpret raw wet-lab data, from gel images and colony plates to microscopy and instrument files.

Which model scored highest on BioBasicsBench?

GPT-6 Astra, with 66.7% of available credit at the xhigh reasoning setting, ahead of GPT-6.1 Sol at 62.0%.

Can other labs run BioBasicsBench?

Not confirmed. The coverage reviewed does not show a public release of the tasks or scoring code, and only OpenAI models were tested.

Sources

1. Measuring how AI agents handle everyday wet-lab science (OpenAI, 6 Nov 2026)
2. OpenAI unveils BioBasicsBench, a benchmark to gauge how well AI agents read raw experimental data (R&D World, 6 Nov 2026)
3. OpenAI Introduces BioBasicsBench to Test AI Agents on Wet-Lab Data (Unite.AI, 6 Nov 2026)