GPT-6-Astra solves 49 of 50 theoretical physics benchmark problems

By Splitfeed EditorialLast updated Oct 15, 2026, 15:31 UTC

On October 15, 2026, researchers published results from the TheoPhysEval benchmark showing that OpenAI's GPT-6-Astra correctly solved 49 of 50 unpublished theoretical physics problems at an accuracy level matching leading human specialists. The cost per problem was $2.40. Notably, at least some of the cases where the model initially appeared to fall short were later found to reflect errors in the human experts' own reference solutions, not in the model's outputs.

The result matters because theoretical physics has historically been treated as a domain where AI progress lagged behind other scientific fields, including mathematics. The $2.40 per-problem cost figure is significant: it suggests that problems requiring specialist-level expertise can now be processed at a price point accessible to almost any research operation. The researchers themselves flag a genuine limitation, noting that the model lacks scientific "taste," meaning the capacity to judge which problems are worth pursuing rather than simply solving problems it is handed. That distinction separates benchmark performance from autonomous scientific contribution. The findings are already prompting discussion about how young physicists should be trained, which signals that practitioners in the field take the results seriously enough to consider structural changes in education.

No published reactions from the Pro-AI, Anti-AI, or Middle Ground camps had appeared at the time of writing. Based on their typical stances, the Pro-AI camp would likely point to the benchmark score and low cost as evidence that AI is approaching genuine scientific partnership with humans. The Anti-AI camp would be expected to stress the "taste" limitation as proof that high accuracy on structured benchmarks does not translate to real scientific creativity or independent discovery. The Middle Ground camp would probably treat both points as valid, framing the result as meaningful but narrow progress that warrants careful institutional planning rather than either alarm or celebration.

The argument about whether benchmark performance translates to autonomous physics research would most directly be tested by whether a model like GPT-6-Astra can identify and frame a novel, consequential physics problem on its own, not just solve one it has been given. Any future evaluation that tests problem selection rather than problem solving would carry significant weight in settling that dispute.