Altman at Dreamforce: an internal model past Astra "can do things that the best mathematicians in the world cannot."

In a session with Marc Benioff uploaded September 16, Altman laid out a ladder: GPT-5.5 "maybe as good as an average math professor," GPT-5.6 "a top one or two percentile math professor," Astra "a little bit better than that," and an unreleased internal model beyond Astra at the top. Unbenchmarked and self-reported, which is the entire counter-argument. It's still the specific ladder OpenAI is now willing to state on stage, which is a different thing from a press release.