OpenAI's LifeSciBench: AI's Real-World Science Impact

OpenAI has released 'LifeSciBench,' a new benchmark test designed to measure how useful AI can be for life science researchers. This test aims to provide evaluations more aligned with actual operations compared to traditional scientific tests. What's interesting is that conventional tests often focus on narrow knowledge or simple question-and-answer formats, which don't accurately reflect real-world capabilities. To address this, OpenAI collaborated with 173 scientists to create 750 tasks across seven categories, covering daily scientific activities. Each task is structured like a scientist asking for help from a knowledgeable colleague. The AI must generate free-response answers after reviewing 1062 attachments, including figures and chemical structure files. The AI's answers are then evaluated on criteria like detail, correct reasoning, and formatting. The results show OpenAI's science-focused AI model, GPT-Rosalind, achieved higher scores than other models, including GPT-5.5, Gemini 3.1 Pro, GPT-5.4, and Grok 4.3. This new benchmark helps us understand AI's practical utility in advancing scientific research.