Artificial intelligence is dramatically reducing the computational cost of solving complex theoretical problems in mathematics, physics and economics, according to new research from the Massachusetts Institute of Technology (MIT).
A study by researchers at MIT FutureTech found that the proportion of advanced theoretical questions successfully answered by AI models increased from just 2.8 percent to 34.6 percent over a 16-month period.
During the same period, the computing cost of producing an answer that received full marks fell from approximately $170 to $15, an elevenfold reduction.
The findings suggest that AI could substantially change how scientific research is conducted, potentially allowing researchers to investigate theoretical questions that previously required considerable time and specialist expertise.
Testing AI against human researchers
The researchers, Gabriel Manso, Claudio Mayrink Verdun and Neil Thompson, evaluated six generations of OpenAI models released between December 2024 and April 2026.
Each model was given the same 254 theoretical questions drawn from recent scientific papers across seven fields, including mathematics, physics and economics.
The models were provided with the questions and their underlying assumptions but were prevented from accessing the original research papers or searching the internet.
Their answers were then assessed against the conclusions reached by the human researchers who wrote the original papers.
The latest model tested achieved full marks on 88 of the 254 questions, compared with just seven for the earliest model.
Improvements were recorded across all seven scientific fields.
What cheaper scientific reasoning could mean
The researchers suggest that falling computational costs could make advanced theoretical methods accessible to scientists working outside their traditional areas of expertise.
It could also encourage researchers to investigate more questions, including those previously considered too expensive or time-consuming to pursue.
However, the study highlights an important limitation: producing a plausible theoretical answer is not the same as establishing new scientific knowledge.
The results were primarily graded using AI, with limited independent mathematical review. The reported costs also exclude human labor and the work required to verify conclusions.
Scientists must still identify worthwhile research questions, test results and determine how discoveries fit into existing knowledge.
The researchers suggest that as theoretical reasoning becomes cheaper, these activities could become increasingly important constraints on scientific progress.
The paper, AI Disruption and Falling Costs in Theory Research, raises the possibility that AI will change not only the speed of scientific discovery, but also how research is organized and which questions scientists choose to investigate.
Main image courtesy of ScienMag.com
