OpenAI's "Mewfour" Is Almost Here: A New Frontier Model Solves Ten Math Puzzles

On August 7, 2026, several outlets reported that OpenAI is preparing to release a new model known internally as "mewfour" and likely to ship under the public name Astra. If the launch happens, it would be the largest model OpenAI has trained since GPT-4.5, marking a clear step back into truly frontier-scale training after a year of more measured updates.

What made the news unusual was the proof, not the press release: an internal version of the model is credited in a recent preprint with ten fresh results in mathematics and theoretical computer science, from sharper upper bounds on high-dimensional sphere packing to a refutation of the Connes rigidity conjecture. At standard Sol API rates, the tokens spent producing those results cost only around two thousand US dollars in total.

The bigger signal is efficiency, not raw size: reaching that level of research output for the price of a domestic flight suggests that the product of model scale and inference efficiency is improving faster than the headline parameter count would suggest. Frontier AI is becoming cheaper to use, even as the models themselves keep getting larger.

The release lands in a tight competitive moment: Anthropic's enterprise-focused models and Google DeepMind's Gemini line have both closed part of the gap, and Chinese labs are scaling quickly. For OpenAI, shipping a credible frontier model now is less about showing off and more about reminding the market who still sets the pace.