OpenAI introduced LifeSciBench, a new benchmark designed to evaluate how well AI systems handle real-world life science research. Unlike existing tests that focus on narrow questions or clean predictions, this benchmark captures the messy, multi-step nature of actual scientific work. LifeSciBench includes 750 expert-authored tasks across seven workflows and biological domains, created by 173 scientists with Ph.D.-level training and industry experience.

Each task mirrors requests scientists make to collaborators, requiring models to interpret evidence, make judgments, and communicate findings with proper caveats. The benchmark incorporates 1,062 artifacts including figures, tables, and chemical files, with 53% of tasks demanding synthesis across multiple data sources. Expert-developed rubrics with 19,020 total criteria grade responses on scientific correctness and operational usefulness, not just final answers.

Validation came from 453 independent reviewers, 97% holding doctorates with an average of 12 years experience. Testing frontier models revealed meaningful progress: GPT-5.5 achieved 25.7% pass rate while GPT-Rosalind reached 36.1%, with notable gains in scientific communication and translation workflows.

LifeSciBench addresses a critical gap in AI evaluation for scientific applications. By grounding tasks in real research practices and requiring nuanced judgment rather than simple fact recall, it provides a more accurate measure of whether AI systems can meaningfully assist life science researchers in drug discovery and related fields.