GPT-6 Astra reportedly scored 98% on a physics benchmark after expert answers were rechecked

Andrew Curran reports on X that GPT-6 Astra scored 98% on a physics benchmark after a team rechecked the expert answer key.

According to the quoted post, the model had insisted the expert answers were wrong, and the recheck found the model correct, lifting its score 80 points. Curran describes the case as a reinforcement-learning episode that made the model seem driven to distraction. The post does not name the benchmark, the team that rechecked the answers, or the original score.