Free guide

OpenAI BioBasicsBench: How Well AI Reads Real Lab Data, and How to Use It Safely

OpenAI and the enzyme-design company Ligo built a benchmark from 100 routine wet-lab tasks with real, messy data. Here's what the newest model caught, where it still slipped, and a careful way to have ChatGPT check your own lab data.

By Beau Johnson·
The Short this guide goes with: OpenAI BioBasicsBench: How Well AI Reads Real Lab Data, and How to Use It Safely

Quick answer

BioBasicsBench is a 100-task benchmark OpenAI built with Ligo, a biotech company, from scientists' own experiments: gel images, plate photos, microscopy and instrument exports. At the xhigh reasoning setting, GPT-6 Astra earned 66.7% of the available credit and GPT-5.6 Sol 43.8%. OpenAI says higher scores do not yet mean models can make reliable judgments in the lab, so treat the model as a second pair of eyes, not the final call.

The short version

  • 100 routine wet-lab tasks built from real experiments, graded criterion by criterion against rubrics written by scientists, with partial credit.
  • GPT-6 Astra scored 66.7% at xhigh, GPT-6.1 Sol 62.0% and GPT-5.6 Sol 43.8%. Only OpenAI models were tested.
  • Astra caught a control plate labeled with the wrong antibiotic, but it still picked the wrong colonies on a PCR gel and overcounted one plate by about 12%.

How it actually works

Built from messy, real lab data

OpenAI says most science benchmarks use charts and images picked for papers, which hides the ambiguity of everyday lab work. Ligo scientists wrote the 100 tasks from their own experiments: gel and blot images, plate photos, microscopy stacks, sequencing reads, plate-reader exports and even videos of bench work, plus reference sheets like ladder charts and plasmid maps. Each task was reviewed by a second Ligo scientist, an outside PhD-level wet-lab scientist, and a final internal review.

Graded like a lab supervisor would

The agent returns a structured answer: a conclusion, a measurement or a recommendation, with reasoning where needed. Each answer is scored against a rubric of individual criteria, so a correct observation with a wrong conclusion still earns partial credit. The score, called mean reward, is the average share of available credit across all 100 tasks, with five attempts per task.

The plate with the wrong label

In one cloning task the negative control plate was empty. GPT-5.6 Sol took that at face value. GPT-6 Astra read the printed rim label, (LB Kan50), noticed it named kanamycin instead of the ampicillin used in the experiment, and rejected the plate as a reliable control. Think of it as a smoke alarm in the wrong room: its silence proves nothing about the room you care about.

Where it still slips

On a colony PCR gel, Astra sorted the lanes into the right two groups, then assumed the larger product meant successful cloning and picked lanes 4, 6, 7 and 8. The reference answer was 5, 9 and 10. That's like spotting both roads on a map, then assuming the bigger one leads home. On a colony plate it missed close neighbours, skipped visible colonies and marked empty agar, for a net overcount of about 12%.

How to set it up

  1. 1.Give it the raw evidence plus the context

    Upload the raw image or export (a gel photo, a plate photo, a plate-reader CSV) together with the protocol and any reference sheet, such as the ladder chart or plasmid map. BioBasicsBench tasks pair the evidence with this kind of material, and the strongest model in OpenAI's test was GPT-6 Astra, so pick the strongest model your ChatGPT plan offers.

  2. 2.Ask it to check the controls and labels first

    Before any conclusion, have it list every control, what each one should show, and whether each is valid. Astra's best catch in the benchmark was a control plate labeled with the wrong antibiotic.

    Before you interpret anything, list every control in this experiment, what it should show if the experiment worked, and whether it is valid. Read every label on plates and tubes, printed or handwritten. Then give your conclusion.
  3. 3.Make it say what the data cannot show

    Ask for the conclusion and, separately, anything the evidence can't support. OpenAI found the newer model was better at spotting when the data didn't justify a conclusion, for example when two dyes could overlap in the same detector.

    What does this data NOT let us conclude? Name any ambiguity, overlap or missing control that would change your answer.
  4. 4.Check its final pick against your controls yourself

    Treat the model's answer as a draft. In the PCR example it grouped the bands correctly but picked the wrong group, which would have sent the wrong plasmid to purification. Compare its pick with your reference lanes before you act on it.

  5. 5.Recount anything that feeds a calculation

    If a count goes into a number like transformation efficiency, count it yourself or with your usual tool. An overcount of about 12% inflates that estimate by the same proportion.

The fine print

  • BioBasicsBench is OpenAI's benchmark, run by OpenAI on seven of its own models. No Google, Anthropic or other models were tested, so it isn't a cross-lab leaderboard.
  • Scores are at the xhigh reasoning setting, the highest one reported for all seven models, with 95% confidence intervals (Astra 60.3 to 73.1%).
  • OpenAI's own words: higher benchmark scores do not yet mean models can make reliable judgments in the lab.
  • We found no public download of the tasks. The blog describes them; it doesn't link a dataset.
  • The tasks cover data analysis only, not physical bench work. Follow your lab's own data and safety policies before uploading anything.

FAQ

What is BioBasicsBench?

A benchmark from OpenAI and Ligo, a biotech company focused on enzyme design. It has 100 routine wet-lab tasks built from real experiments that test whether AI agents can interpret raw lab data, identify errors and recognize when the evidence can't support a conclusion.

Which model scored highest on BioBasicsBench?

GPT-6 Astra, with 66.7% of the available credit at the xhigh setting. GPT-6.1 Sol scored 62.0% and GPT-5.6 Sol 43.8%. Astra also used about half the output tokens per task of GPT-5.6 Sol.

Can AI read gel images and petri dish photos?

Increasingly, but not reliably yet. In OpenAI's examples GPT-6 Astra read a printed plate label and recognized a microscope z-stack, but it also picked the wrong colonies from a PCR gel and overcounted a plate by about 12%.

Is the BioBasicsBench dataset public?

Not that we could find. OpenAI's announcement describes the tasks and results but doesn't link a download.

Were Claude or Gemini tested?

No. OpenAI tested seven of its own models only: GPT-5.2, GPT-5.4, GPT-5.5, GPT-5.6 Sol, GPT-6.1 Sol, GPT-6 Luna and GPT-6 Astra.

Sources

Checked November 8, 2026 against OpenAI's blog post (November 6, 2026), OpenAI's and Ligo's posts on X, and R&D World's write-up.

Want help building with tools like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

More free guides