OpenAI introduced a new test of how artificial intelligence handles everyday laboratory evidence, reporting stronger results from its GPT-6 Astra model while warning that higher scores do not yet mean reliable judgments in the lab.
The benchmark, developed with enzyme-design company Ligo, is called BioBasicsBench. It tests AI agents on 100 tasks drawn from routine laboratory work, including interpreting experimental images, spotting errors and recognizing when the evidence cannot support a conclusion.
For researchers and biotechnology businesses, the issue is practical: a mistaken reading of an experiment or a missed problem with its controls can send scientists toward the wrong conclusion and waste time on follow-up work.
In its November 6 announcement, OpenAI said GPT-6 Astra earned an average of 66.7% of the available scoring credit at the test’s “xhigh” reasoning setting, compared with 43.8% for GPT-5.6 Sol. Those figures are average benchmark scores, not the percentage of experiments completed without mistakes.
Testing raw evidence, not polished results
The tasks use material such as gel images, microscope images, sequencing files, instrument readouts and videos of laboratory work. Scientists built the questions from their own experimental data and paired them with expert-written grading criteria.
OpenAI said the evaluation covered seven models and 37 combinations of models and reasoning settings. Each combination made five attempts on each of the 100 tasks, producing 18,500 results.
The company also documented continuing mistakes. In one example, Astra selected the wrong colonies from experimental results after an assumption overrode the comparison supported by the controls. In another, it overcounted colonies by approximately 12% relative to the benchmark reference.
OpenAI said the benchmark is intended to track progress in interpreting raw data and checking whether conclusions are supported. The results do not establish that an AI system can replace a scientist’s review.
JBizNews Desk | Technology
© JBizNews.com. All rights reserved. Unauthorized reproduction or redistribution is strictly prohibited.
