Altman says internal model beyond Astra surpasses top mathematicians

At Dreamforce 2026, Sam Altman laid out OpenAI's math capability ladder in a conversation with Salesforce CEO Marc Benioff, and disclosed a stronger internal model beyond Astra. He described an unusually steep takeoff in mathematical capability, while noting that measuring these models by human standards is becoming less meaningful. The claims remain unverified.

Confirmed

• Altman said GPT-5.5 performs roughly at the level of an average math professor.
• GPT-5.6 was positioned at the top 1-2% of professors.
• He claimed an internal model codenamed past Astra can already do things even the world's best mathematicians cannot.
• The full conversation video was posted by @rohanpaulai; @acoolrandomusername, @haider1, @gerardsans, @perksverse and others relayed consistent accounts of the capability ladder.

Unconfirmed

• These remain verbal claims with no public benchmarks or independent verification. VraserX noted that if true, the model should be handed to researchers quickly to accelerate scientific discovery; nikolamr64990 flagged the account as third-party relay, not officially confirmed.

Why it matters

• If the internal model truly has this capability, AI may have surpassed top human levels in frontier mathematics research, with deep implications for science; but the actual capability boundary awaits public evidence.

2026-09-16 ~ 2026-09-18 · 9 related posts

• Altman says internal post-Astra OpenAI model can solve problems the world's best mathematicians cannot — rohanpaul_ai · 2026-09-16
• Sam Altman claims OpenAI has an internal model solving problems top mathematicians can't — VraserX · 2026-09-16
• Altman: GPT 5.5 Matches an Average Math Professor, Internal Model Beats World's Best — acoolrandomusername · 2026-09-16

6 near-duplicate retellings: rohanpaul_ai · haider1 · haider1 · perksverse · nikola_mr64990 · gerardsans