OpenAI's Internal Model Solves Decade-Old Math Problems for Under $2,000 Each

Simon Willison · rss · 2026-08-02

Following Anthropic's recent cryptography work, OpenAI demonstrated the mathematical prowess of its next-generation internal model (codenamed Astra / GPT-5.6). They tasked the model with solving ten mathematical problems that had seen no progress for at least a decade, claiming the inference token cost was less than $2,000 each.

Demonstrating high transparency, OpenAI released an accompanying paper and open-sourced the Lean 4 formalizations of their results on GitHub. This breakthrough has caused a collective "Deep Blue" moment among mathematicians, echoing Terence Tao's vision of a "big mathematics" era characterized by massive human-AI collaboration.

Original post →

More from AGI Musings

AGI Musings channel →