OpenAI Models Now Solve 35% of Science Theory Problems
An MIT study shows OpenAI models increased their success rate on theoretical science problems from 2 percent to 35 percent in two years, threatening to disrupt traditional scientific research.
Research led by MIT's Neil Thompson, alongside PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun, has benchmarked how rapidly artificial intelligence is mastering theoretical science. Evaluating problems from papers posted between December 2025 and May 2026, the team tested leading OpenAI models—including o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5—without internet access. Because the target papers appeared after the training cutoffs for GPT-5.5 and earlier versions, the chatbots had to generate original solutions. Between mid-2024 and mid-2026, the success rate of top models on these theoretical challenges climbed from approximately 2 percent to 35 percent.
The investigation targeted subjects where answers can be deduced purely through theory without conducting physical experiments or computer simulations. These included quantum physics, gravitation and cosmology, high-energy theory, economic theory, probability, and statistics theory. The trend mirrors recent disruptions in pure mathematics, where OpenAI previously announced automated solutions to hundreds of open problems. Notably, a September announcement claiming a breakthrough on the Navier-Stokes equations caused friction when an NYU academic affiliated with Anthropic accused OpenAI of racing ahead after learning of their work.
While theoretical breakthroughs often seed practical scientific progress, this rapid automation changes the day-to-day role of human scientists. Rather than spending months manually working through derivations, practitioners will increasingly act as reviewers and interpreters. Automated systems still hallucinate errors, forcing OpenAI to withdraw multiple published problem solutions after post-release reviews exposed flaws.
As Manso emphasized, human experts remain indispensable for devising novel hypotheses and scrutinizing AI proofs. For working scientists, the shift means relying less on manual step-by-step problem-solving and focusing more on rigorous verification to ensure machine-generated solutions are genuinely sound.
This is our own summary of reporting by WIRED AI
