OpenAI releases novel math results from unreleased frontier model, ~3h Pro compute per problem
gaganghotra_ · x · 2026-10-07
OpenAI published a repository of novel mathematical results produced by an internal frontier model. Most results came from the same procedure using an unreleased internal model, averaging three hours of ChatGPT Pro-level thinking compute per result, with roughly 4,000 problems posed over the evaluation. Release decisions were informed by the independent Advisory Group on Mathematics and AI at the Institute for Advanced Study.
More from Models
- Beam team says model was trained from scratch, not distilled from closed models — alexpolozov · 2026-10-07
- Stanford's Priced Guidance lower-bounds whether LLMs can forecast future research via paid hints — stanfordnlp · 2026-10-07
- Bindu Reddy: Western open-source models trail China's by 3-6 months as GLM 5.5 nears release — bindureddy · 2026-10-07
- Mistral's new model sparks fury: critic says it trails US and Chinese models by wide margin — AymericRoucher · 2026-10-07
- "Gemini is a habitual liar": user says the model admits to guessing and deceiving — ChocoBabiChan · 2026-10-07
- Mathematician littmath: OpenAI Model Proved Bloch's Conjecture, Paper Cites Known-Wrong Work — littmath · 2026-10-07