OpenAI releases LifeSciBench: an expert-written + expert-reviewed life science AI benchmark

OpenAI releases LifeSciBench, a life sciences AI benchmark written and reviewed by domain experts. LifeSciBench contains 750 assessment tasks extracted from real scientific research activities, covering areas such as molecular biology, biochemistry, genetics, and drug discovery.

LifeSciBench is unique in its assessment methodology: each task is accompanied by a fine-grained scoring rubric developed by a team of expert reviewers, totaling 19,020 scoring points. This ensures the objectivity and accuracy of the evaluation and avoids the common problem of "brushing the rankings" in traditional benchmark tests.

OpenAI also released GPT-Rosalind, a life science-specific model version based on GPT-5.5, which achieved leading results on LifeSciBench. OpenAI said LifeSciBench will be open to academia to promote standardized assessment of AI in the life sciences.

The launch of LifeSciBench fills the gap in standardized assessment of AI in the life sciences. The design idea that all 750 tasks come from real scientific research activities is worthy of reference by domestic AI evaluation institutions - not artificially constructed "exam questions", but "practical questions" from the front line. This means that models evaluated through LifeSciBench will perform more predictably in real scientific research scenarios.