Wet-lab AI has been passing tests that look nothing like wet-lab work.
OpenAI says its new BioBasicsBench finally grades agents on the mess, with 100 routine tasks built with Ligo from raw gels, plates, scope stacks and instrument exports. At the xhigh reasoning setting, GPT-6 Astra earned 66.7% of available credit, compared with 43.8% for GPT-5.6 Sol, according to the company.
I read that jump as progress, not readiness.
The strongest evidence is control handling, the place where lab mistakes turn expensive. In one cloning task GPT-5.6 Sol accepted an empty negative control plate at face value, while Astra spotted the rim label reading LB Kan50 and rejected it because the experiment used ampicillin, according to OpenAI. That is exactly the kind of bench detail paper-figure benchmarks never test, and trade coverage from R&D World picked up the same example as proof the test targets raw data reading.
The second gain is restraint. In a yeast viability task with overlapping SYTOX and FUN-1 detection at 525 to 540 nm, Sol called cells dead on brightness alone while Astra flagged the spillover and asked for better controls, according to the company. Astra also read disappearing cells in a time series correctly as a z-stack through depth, not biology. Even inside OpenAI's lineup the gap is wide, with Astra at 66.7% clearing GPT-6.1 Sol at 62.0% and GPT-6 Luna at 38.3%, per the reported means.
The concession is that OpenAI shows Astra failing in ways a second-year would catch.
It groups colony PCR bands correctly then picks the wrong group, naming lanes 4, 6, 7 and 8 instead of the reference 5, 9 and 10, which would send the wrong plasmid to purification. It generalizes that vacuoles look darker while ignoring lighter vacuoles in the same field that break the rule with focus change. It overcounts one transformation plate by about 12% by merging neighbors, missing visible colonies and marking empty agar, which directly inflates the transformation efficiency estimate. Those are not edge cases. They are counting, controls and contrast.
My judgment survives that record because the benchmark exposes it instead of hiding it behind curated charts. If you are building lab agents, grade them on raw plates and failed controls before you pitch reliability.
