A theoretical paper in mathematics or a related quantitative science represents a costly investment of expert time. AI has the potential to revolutionize this type of knowledge production, but early models singularly failed to do so. Frontier models, today, however are claiming important discoveries, but some of these claims are contested. To systematically assess how AI’s capabilities are changing, we compare how well model releases from December 2024 to April 2026 solve the cutting-edge theory questions by using cases where we have ground-truth human answers that were unavailable to AI models when answering. These answers come from 254 recent arXiv papers from seven scientific domains spanning mathematics, physics, and economics. We show that progress has been enormous. Over these sixteen months, the share of problems for which models reached the same conclusion as the human authors rose from 2.8% (o1) to 34.6% (GPT-5.5), an increase of 22 percentage points per year. Gains reach every subfield at similar rates. These results were achieved at minimal cost, approximately $5 per attempt or ($15 per success) for GPT-5.5. We interpret these results as signs of a rapidly evolving shift in science in which some areas become substantially more cost-effective, making powerful theoretical tools available to a wider range of scientists.