GPT-6 Astra scores 88% on Serafim's reasoning-induction benchmark

Derya Unutmaz, MD @DeryaTR_Originally posted 2026.09.06

Open original on X

This is an unbelievable and shocking result! GPT-6 Astra got 88% correct on Serafim’s extremely challenging reasoning-induction benchmark, which is now almost saturated! This benchmark was extraordinarily difficult for previous models, including Fable 5.1, which got only 33% correct. Astra is also returning simple, highly generalizable hypotheses, in striking contrast to previous models. Strong support for my declaration that GPT-6 Astra is unequivocally AGI!

This post quotes another post: x.com/s_batzoglou/status/2096407011885986187