OpenAI released LifeSciBench, a benchmark designed to measure how well AI systems can support realistic life science research tasks rather than only answer isolated biology questions. According to OpenAI, the suite was created and reviewed by practicing life scientists with industry experience and Ph.D.-level training.
Benchmark design and scope
LifeSciBench contains 750 expert-authored tasks spanning seven workflows: evidence handling, analysis, design and optimization, scientific reasoning, validation and operations, translation, and scientific communication. The dataset includes 1,062 attached artifacts (figures, PDFs, tables, sequence and structure files, and web references), and OpenAI says 53% of tasks require models to interpret or synthesize information from at least one artifact.
Task authorship involved 173 scientists; grading is driven by 19,020 rubric criteria (an average of 25 criteria per task). OpenAI reports that accepted tasks averaged six automated review cycles and at least two rounds of expert review. Overall, 79% of tasks require multiple reasoning or decision-making steps, with an average of four steps per task.
Validation and reported results
OpenAI describes an independent validation by 453 reviewers who were not involved in task authorship. The company reports that 97% of reviewers held a Ph.D. or equivalent, with an average of 12 years of experience and 14 peer-reviewed publications; agreement on task quality exceeded 96% across assessment categories.
OpenAI also reports benchmark results comparing frontier models. The overall pass rate improved from 25.7% for GPT‑5.5 to 36.1% for GPT‑Rosalind. The largest reported gains were in Scientific Communication (56.3% to 71.1%, n=9) and Translation (36.8% to 57.7%). On rubric-level measures, OpenAI reports GPT‑Rosalind scored 44.7% on tasks requiring actionable outputs versus 29.1% for GPT‑5.5, and 44.8% on uncertainty-and-caveat handling versus 29.3%.
Performance remained weaker on artifact-heavy and design-heavy tasks. Reported pass rates for GPT‑Rosalind were 30.7% in Design, Optimization & Prediction and 30.3% in Analysis. Pass rates for GPT‑Rosalind fell from 45.1% on text-only tasks to 28.1% when artifacts or URLs were present (GPT‑5.5 fell from 29.9% to 21.9%). Exact-output tasks were also challenging: OpenAI reports 14.8% pass on numeric tasks, 24.0% on sequence or structure outputs, and 27.3% on construct-generation tasks.
OpenAI notes that models sometimes earn substantial partial credit: for GPT‑Rosalind, 109 tasks had pass rates below 20% while earning at least 50% of rubric reward.
OpenAI frames LifeSciBench as a step toward evaluating task-level capability and says the next step is deployment studies to measure impact in live research settings. The benchmark was published June 17, 2026, and OpenAI invites contributors and requests for access to GPT‑Rosalind.
Original source: OpenAI News