OpenAI unveils 100 experimental tasks… Astra's average score 66.7%

OpenAI unveiled BioBasicsBench, a 100-task benchmark evaluating AI agents' ability to interpret raw wet-lab data from experiment photos and instrument outputs and to flag errors or weakly supported conclusions. OpenAI reported that GPT-6 Astra averaged 66.7% under its top xhigh reasoning setting, while GPT-5.6 Sol scored 43.8%. Both figures are averages of per-task rubric scores, not accuracy or success rates, so 66.7% does not mean 66.7 of 100 tasks were solved. The benchmark matters because it targets data-reading and judgment skills rather than end-to-end research outcomes or lab automation success. What to watch next is whether rivals adopt the benchmark and how scores shift across reasoning settings and future model releases.