MIT FutureTech: AI theory answers jump to 34.6%, cost $15

The MIT FutureTech study puts numbers on a shift that has been hard to pin down: how fast AI is getting at theoretical reasoning, and at what price. Six OpenAI model generations released between December 2024 and April 2026 were given identical questions, with the original papers and the internet withheld — a setup that tests reasoning rather than retrieval. Full marks went from 7 of 254 questions to 88, and the computing cost of a full-marks answer dropped roughly elevenfold, from about $170 to $15.

The authors — Gabriel Manso, Claudio Mayrink Verdun and Neil Thompson — frame the consequence in terms of who can afford to ask. Falling costs could open advanced theoretical methods to scientists working outside their traditional areas of expertise, and could make previously prohibitive questions worth pursuing. The gains were not confined to one discipline; all seven fields improved.

The paper is careful about what this does not show. Producing a plausible theoretical answer is not the same as establishing new scientific knowledge, grading leaned heavily on AI with limited independent mathematical review, and the cost figures exclude human labor and verification. The researchers suggest that as reasoning gets cheaper, identifying worthwhile questions, testing results and fitting findings into existing knowledge could become the tighter constraints on progress — a shift in how research is organized, not just how fast it moves.