OpenAI BioBasicsBench: GPT-6 Astra Scores 66.7% on Lab Work

§ 01What BioBasicsBench tests#

Most science benchmarks use figures that were cleaned up for a paper. BioBasicsBench uses the messy originals: gel and blot images, plate photos, microscopy stacks, sequencing reads, plate-reader exports and even videos of bench work. Scientists built each task from data from their own experiments and wrote a grading rubric for it. Every task then passed three reviews: a Ligo scientist who had not written it, an outside PhD-level scientist working in a wet lab, and a final internal check.

Answers earn partial credit per rubric criterion, so a correct conclusion with a wrong measurement scores differently from a wrong conclusion. Nineteen tasks specifically test whether the agent can say the evidence is not enough.

One task can test several skills, so the skill columns do not add up to the task total. The most common inputs are gels and blots (31 tasks), protocols and lab notes (27) and spreadsheet or instrument exports (20).

§ 02The results#

OpenAI ran seven of its own models across 37 model and reasoning-effort settings, five attempts per task, for 18,500 graded answers. At xhigh, the highest setting all seven share:

Three things stand out in that table:

• The jump from GPT-5.6 Sol is real. Astra’s interval (60.3-73.1%) does not overlap Sol’s (37.6-50.0%).
• Astra and GPT-6.1 Sol are close. Their intervals overlap heavily, so the 4.7-point gap at the top is not a clear win.
• Better answers came from less writing. Astra used 4,704 output tokens per task, about half of GPT-5.6 Sol’s 9,099 and a sixth of GPT-5.2’s 27,620. OpenAI’s reading is that “stronger native multimodal capabilities may be key to improving performance on everyday wet-lab analysis.” In other words, seeing the image better beat thinking about it longer.

§ 03Where Astra got better#

OpenAI walks through three cases where GPT-5.6 Sol went wrong and Astra did not.

• It knew what the data could not show. Sol called cells dead because they glowed in the channel for a dead-cell dye. But that channel (470-540 nm) overlaps the window of a second dye (525-580 nm) at 525-540 nm, so brightness alone proved nothing. Astra flagged the overlap and asked for better controls.
• It read the label on the plate. In a cloning experiment, Sol saw no colonies on the negative-control plate and treated the experiment as clean. Astra noticed the rim label read “LB Kan50”, a kanamycin plate, while the experiment used ampicillin, and rejected the plate as a valid control.
• It kept the context in mind. Shown cells fading in and out across microscope frames, Sol looked for a biological reason. Astra recognized a z-stack, the same cells imaged at different depths.

§ 04Where it still fails#

The failure cases are the most instructive part of the release, because each one is a mistake a careful scientist would not make.

• An assumption beat the controls. On a colony PCR gel, Astra sorted the bands into the right two groups but assumed the larger product meant successful cloning. It picked lanes 4, 6, 7 and 8; the controls point to 5, 9 and 10. Following that answer would mean purifying the wrong plasmid.
• It discarded evidence that did not fit. Astra concluded that vacuoles look darker than the surrounding cell, even though lighter vacuoles were visible in the same image and it had itself noted that contrast changes with focus.
• It miscounted a plate. Astra merged close neighbours, missed visible colonies and marked empty agar, for a net overcount of about 12%. Because transformation efficiency is calculated from that count, the efficiency estimate is inflated by the same 12%.

§ 05What it means beyond the lab#

BioBasicsBench is a biology test, but the pattern applies to any agent that reads real-world evidence at work: invoices, dashboards, contracts, screenshots. The costly error is rarely a wrong fact. It is a confident conclusion built on a check that was never valid, such as the wrong control plate, a mislabeled column or a stale field. A benchmark that grades “recognize when the evidence cannot support a conclusion” as its own skill is measuring the right thing, and even the best model here earns two thirds of the credit.

OpenAI also notes the dual-use side: “Tracking these capabilities will also help us understand their implications for biosafety and biosecurity.”

§ 06Where CellCog fits#

Our conflict, declared: we build CellCog, an AI employee platform, and our agents run on Anthropic’s Claude Opus 5.5, which was not part of this benchmark. A CellCog AI employee can read the same kinds of inputs, from images and PDFs to CSV and spreadsheet exports, and keep a running record of its work in its own workspace and task board. The lesson from BioBasicsBench is how to use one: let the agent do the first read and draft the analysis, and keep the judgment call on controls and conclusions with the scientist.

§ 07What is not established yet#

• Other labs’ models. OpenAI tested only its own seven models. There are no Claude, Gemini or open-weight scores on BioBasicsBench yet.
• Independent replication. The announcement does not say whether the 100 tasks will be released for others to run, and Ligo, the co-developer, also supplied the data.
• Real lab outcomes. A higher rubric score is not evidence that the agent’s calls improve results at the bench; OpenAI says as much itself.

§ 08Where to watch#

• Whether OpenAI or Ligo publish the task set so other labs can be scored.
• The first non-OpenAI results, especially Claude and Gemini models with strong image reading.
• Follow-up work on the failure modes: control checking, colony counting and contradicting evidence.

§ 09Sources#

• Measuring how AI agents handle everyday wet-lab science, OpenAI, November 6, 2026.
• OpenAI news feed, publication date for the BioBasicsBench post.
• OpenAI’s announcement on X, November 6, 2026.