BioBasicsBench tests wet-lab analysis, and the best model earns 66.7%

OpenAI and enzyme-design company Ligo have introduced BioBasicsBench, 100 tasks drawn from routine wet-lab work. The best tested model, GPT-6 Astra, earned 66.7% of available credit. Its documented mistakes could still send a scientist to the wrong sample or distort an efficiency calculation.

One hundred tasks begin with raw evidence rather than publication figures

The benchmark covers five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods and molecular cloning. Inputs include gels, colony plates, microscopy, sequencing reads, instrument exports and protocol video.

Scientists authored tasks from their own experiments and created scoring rubrics. Each received review from another Ligo scientist, an external PhD-level wet-lab scientist and a final internal pass. Seven models across 37 model and reasoning-effort combinations attempted every task five times, producing 18,500 responses.

The score rewards recognizing when evidence is insufficient

At xhigh reasoning effort, GPT-6 Astra averaged 66.7% with a 95% confidence interval of 60.3 to 73.1%. GPT-5.6 Sol scored 43.8%. Credit covers more than the final number. Rubrics assess measurements, reasoning, controls and the ability to reject conclusions unsupported by evidence.

That distinction matters more than another biology knowledge quiz. Routine laboratory work contains blurred images, invalid controls and conflicting observations. A useful agent must recognize when to stop calculating and request another experiment.

Two-thirds of the credit still permits expensive errors

OpenAI shows Astra correctly identifying fluorescence spillover and an invalid control plate. It also selected the wrong colonies in a colony PCR task. In another example, it overcounted colonies by about 12%, which would inflate the estimated transformation efficiency by the same proportion if volume and dilution remained fixed.

The results are reported by OpenAI and compare only OpenAI models. The public page does not link a dataset, code or paper, so independent teams cannot yet audit task selection and scoring.

Open data and tests in other laboratories will determine its value

The first useful signal is release of tasks, rubrics and suitably anonymized raw data. Without them, outsiders can only watch a vendor score its own models. The second is performance in laboratories beyond Ligo, with different instruments, protocols and failure patterns.

Deployment decisions should focus on high-risk subsets rather than the 66.7% average. A laboratory needs evidence that the agent checks controls, calibration and uncertainty precisely where an error changes the next experiment.

Lilith's verdict

An agent earning two-thirds of the credit may sort images well. Keep it away from the pipette until it spots the bad control before the scientist who is about to lose a week to it.