Leaked chart analysis: Millennium Prize solved by GPT-6.2 Astra, internal model ahead

YashasGunderia · x · 2026-11-23

Unverified interpretation of OpenAI's pretraining progress chart: the model that solved the Millennium Prize problem was likely GPT-6.2 Astra, whose level aligns with that difficulty line. OpenAI's internal model is reportedly another 0.03 lower in perplexity—roughly the GPT-5 to GPT-5.4 jump—and per Pachocki, smaller perplexity gains mean bigger intelligence leaps closer to the floor.

Related event: OpenAI Discloses Pretraining Perplexity Data, Proposes Industry Safety Threshold(45 posts)→

Original post →