The questions had human-authored ground-truth answers that were unavailable to the models when they answered. The share of problems on which a model reached the same conclusion as the authors rose from 2.8% for o1 to 34.6% for GPT-5.5. The researchers estimate approximately $5 per attempt and $15 per successful match for GPT-5.5. The result measures answer agreement on this test set; it does not establish independent discovery or validate a complete AI-led research process. Source: https://futuretech.mit.edu/publication/ai-disruption-and-falling-costs-in-theory-research
The use of AI in science raises questions on several levels.
In one minute
MIT FutureTech researchers compared models released from December 2024 through April 2026 on 254 cutting-edge theory questions drawn from recent arXiv papers across seven domains in mathematics, physics, and economics.
What changed: Adds MIT CSAIL findings that frontier models’ success on theory questions rose from 2.8% to 34.6% in 16 months, at roughly $15 per successful answer.