Seriously, It’s AI Seriously, It’s AIAI News • Tools • Trends Subscribe
Latest
AI News

OpenAI's AI Just Took a Real Wet-Lab Exam — GPT-6 Astra Scored 66.7%, and the Mistakes Matter More

November 07, 2026 By Seriously, It’s AI 0 Comments


 I spend a lot of time reading AI benchmarks, and most of them test the same tidy things: multiple-choice questions, clean charts, textbook problems. OpenAI's new test is different. On November 6, the company introduced BioBasicsBench, a benchmark built with enzyme-design firm Ligo that hands AI agents the raw, messy evidence of real laboratory work — gel images, colony plates, microscope stacks, sequencing reads — and asks them to reach conclusions a bench scientist would accept. GPT-6 Astra, OpenAI's newest model, earned 66.7% of the available credit. That number is interesting. What it hides is more interesting.

What BioBasicsBench actually tests

BioBasicsBench is 100 tasks drawn from real experiments, spread across five areas of everyday wet-lab work: protein methods, imaging and cytometry, assays and quantification, nucleic acid methods, and molecular cloning. Each task comes with an expert-written grading rubric, and every task passed review by a Ligo scientist, an external PhD-level wet-lab scientist, and a final internal check. The point, OpenAI says, is that everyday lab work is full of failed experiments, ambiguous images, conflicting observations, and troubleshooting — the stuff paper-perfect benchmarks never show. The agents have to interpret the evidence, spot errors, and say when the evidence cannot justify a conclusion.

The scores — and what 66.7% really means

At the highest reasoning setting OpenAI reported for all seven models tested ("xhigh"), GPT-6 Astra earned a mean reward of 66.7%, with a 95% confidence interval of 60.3–73.1%. GPT-5.6 Sol, the previous generation, scored 43.8%. GPT-6.1 Sol landed at 62.0%. In total OpenAI ran 37 model-and-effort combinations with five attempts on every task — 18,500 graded answers. Read the number carefully, though. It is a mean rubric reward — partial credit across findings, measurements, reasoning, and knowing when to say "not enough evidence." It is not a pass rate. Astra did not solve two-thirds of the tasks outright, and a model can collect most of a task's credit while still reaching the wrong conclusion. The gap between generations is widening fast. The same week, Tencent's Hy4 arrived with vision and computer use for the first time — another sign that the frontier is moving from answering questions to doing work. Read our take on that launch here :Tencent Hy4 Is Here: Vision and Computer Use Arrive, but the Weights Are Still Coming 

Where Astra still got it wrong

OpenAI's own write-up is refreshingly candid about the failures. In a colony-PCR task, Astra picked lanes 4, 6, 7 and 8, assuming a larger PCR product meant successful cloning; the reference interpretation picked lanes 5, 9 and 10. In a colony-counting task, it failed to separate close neighbours, missed visible colonies, and marked empty agar — overcounting by about 12% against the reference. These are exactly the mistakes that would send real lab work down the wrong path, and OpenAI says plainly that higher scores do not yet amount to reliable laboratory judgment. I checked the coverage myself: the failures are in OpenAI's own announcement, not something a critic dug up afterwards.

Why this matters

A model that can read a gel is a model that can actually help in a lab, not just talk about labs. If benchmarks like this become standard, the next generation of lab assistants gets measured on evidenceinstead of eloquence — and that is a genuine step toward AI doing science rather than summarising it. OpenAI has had a busy week: it also shipped Images 2.6, a 50% faster image model with animated GIFs — though the real cost per image quietly rose. Read that story here: OpenAI Images 2.6 Is Here: 50% Faster, With GIFs — but Watch the Real Cost But keep the limits in view. Every model tested was OpenAI's own, no human baseline was published, and at launch there was no public link to the dataset, the rubrics, or the scoring code — so outside labs cannot yet verify the scores or run their own models against the same tasks. Until independent checks exist, BioBasicsBench measures OpenAI against OpenAI. The company is pushing its models into real-world institutions at speed — from Abu Dhabi's machine-readable government to the lab bench — and the scrutiny has to keep up. Read about the Abu Dhabi launch here:  Abu Dhabi and OpenAI Launch a Government AI Can Read

What to watch next

Three things. First, whether OpenAI publishes the dataset and rubrics so others can reproduce the scores. Second, whether a human baseline appears, so we know what 66.7% means next to a trained scientist. Third, whether labs start using this kind of test to decide which models they trust near real experiments. If the dataset stays private, the benchmark is a press release with numbers. If it opens up, it could become the standard lab exam for AI. BioBasicsBench is the most honest kind of benchmark — one built to expose failure, not to crown a winner. Astra leads, but the lead comes with an asterisk the size of a petri dish. Would you trust an AI to read your lab results yet — or is 66.7% still a "check its homework" score? Let us know in the comments.

Sources

Comments

Browse by Category

Stay Ahead with Seriously, It’s AI

Get the latest AI news, trends and expert insights — explained simply, every day.

Follow Us →

Stay updated with AI

One short email a week. No spam, unsubscribe anytime.