GPT-6 Astra solves 49 of 50 problems on TheoPhysEval physics benchmark

Techub News reported that GPT-6 Astra solved 49 out of 50 unpublished theoretical physics calculation problems in the TheoPhysEval benchmark, with accuracy described as comparable to that of top experts. Researchers said some cases that initially appeared to be failures led to a different finding: the experts’ own reference solutions were incorrect. According to The Decoder, the team said physics has not yet seen a breakthrough comparable to what has happened in mathematics because AI agents still lack "taste." Even so, they added that autonomous, top-tier AI research may arrive soon. The report centers on benchmark performance and the researchers’ stated view of where AI research in physics stands now.