OpenAI’s GPT-6 Astra Scores 66.7% on New Wet-Lab Test as Errors Persist

OpenAI on Friday, Nov. 6, published BioBasicsBench, a 100-task test of whether AI agents can read the raw evidence of routine biology lab work. These are gel images, colony plates, microscopy and instrument exports, not the polished figures found in papers. On the OpenAI BioBasicsBench results, its newest model, GPT-6 Astra, earned an average of 66.7% of the available credit at the company’s “xhigh” reasoning setting. GPT-5.6 Sol scored 43.8%, according to OpenAI.

The company built the benchmark with Ligo, a biotech firm focused on enzyme design. OpenAI is candid that a higher score is not the same as a reliable lab assistant. The same post documents Astra picking the wrong colonies, discounting contradictory images and overcounting a plate by about 12%.

What BioBasicsBench measures

The 100 tasks span five areas of everyday wet-lab work: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Scientists wrote each task from their own experiments and paired it with an expert grading rubric. Each one then went through three reviews: by a Ligo scientist who had not written it, by an outside PhD-level wet-lab scientist, and a final internal check.

The inputs are deliberately messy. By OpenAI’s count, 31 tasks include gels and blots, 27 include protocols or handwritten notes, 20 include CSV or instrument exports, 14 include raw microscopy and 12 include sequencing files. Seven include videos of bench work. Answers earn partial credit against individual rubric points, so a correct conclusion reached with flawed reasoning scores lower than a fully correct analysis.

OpenAI tested seven models across 37 model-and-effort settings, with five attempts on each task per setting, for 18,500 graded results.

By the numbers: a 22.9-point jump, with wide error bars

At xhigh effort, GPT-6 Astra’s 66.7% beat GPT-6.1 Sol (62.0%) and GPT-5.6 Sol (43.8%). Older GPT-5 models ranged from 31.2% to 35.4%. Astra’s lead over GPT-5.6 Sol is 22.9 percentage points, a gain of about 52% (TBR calculation from OpenAI’s published table). It also used about half the output: 4,704 tokens per task versus 9,099, a drop of roughly 48% (TBR calculation).

The uncertainty is large. OpenAI’s 95% confidence interval for Astra runs from 60.3% to 73.1%, and GPT-6.1 Sol’s runs from 55.6% to 68.4%. The two newest models’ ranges overlap, so the benchmark does not cleanly separate them.

Where GPT-6 Astra improved

OpenAI points to three kinds of gains over GPT-5.6 Sol, and it credits stronger native multimodal ability for much of the improvement:

• Knowing what the data cannot show. In a cell-viability task, GPT-5.6 Sol called cells dead based on one dye’s signal. Astra flagged that a second dye’s light overlaps the same detector window and asked for better controls.
• Reading fine detail. In a cloning task, Astra noticed that a “negative control” plate was labeled “LB Kan50,” a different antibiotic from the ampicillin used, and rejected it as a control. GPT-5.6 Sol relied on it.
• Holding context. Shown cells drifting in and out of focus, Astra recognized a z-stack, a series of images taken at different depths. The older model hunted for a biological explanation.

Where it still fails

OpenAI also published the failures. In one colony PCR task, Astra grouped the gel bands correctly but attached the identities backward. It proposed lanes 4, 6, 7 and 8 when the reference answer is lanes 5, 9 and 10, a mistake that would have sent a scientist to purify the wrong plasmid. In a microscopy task, it called vacuoles “generally darker” even after inspecting lighter ones in the same image. On a colony plate, it merged neighbors, missed colonies and counted empty agar, ending about 12% high. With the same plated volume and dilution, that error inflates a transformation-efficiency estimate by the same share, OpenAI notes.

The post ends with a caution: “Higher benchmark scores do not yet mean models can make reliable judgments in the lab.” OpenAI adds that tracking these capabilities also informs its biosafety and biosecurity work.

Why it matters

The release continues a busy week of capability disclosures from the company. It follows Thursday’s GPT-6 Cyber security model results and, on the same day, a security partnership with Hacktron to test how OpenAI’s own systems connect to third-party services. For labs, the takeaway is narrower: the benchmark measures analysis of evidence a scientist has already produced, not autonomous experimentation. It is OpenAI’s own test, built with a partner, and outside groups have not yet reproduced the scores. Readers comparing labs can find context in our OpenAI company profile and our guide to what AI benchmarks actually measure.

Sources

• OpenAI — Measuring how AI agents handle everyday wet-lab science (BioBasicsBench), Nov. 6, 2026
• OpenAI — Partnering with Hacktron to find vulnerabilities in OpenAI’s systems, Nov. 6, 2026
• OpenAI — GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents, Nov. 5, 2026
• The Verge — OpenAI is stepping back from NDAs for data center projects