OpenAI Astra solves 10 open math problems ($2K compute); Fields Medalist endorses rigor

OpenAI announced on August 1, 2026 that an internal, unreleased version of Astra, its next major model family, solved ten previously open problems in mathematics and theoretical computer science, publishing formal Lean proofs on GitHub for independent verification. Problems include establishing the existence of non-sofic groups (a central question in group theory), new upper bounds on sphere-packing density down to the Cohn-Elkies threshold, and other frontier mathematics. The entire research effort consumed roughly $2,000 in compute. Fields Medalist Timothy Gowers examined the proofs and stated he would recommend at least one for publication in Annals of Mathematics without hesitation.

The novelty lies in both the achievement and the verification method. Unlike benchmark claims, which can be gamed or saturated, these results are independently verifiable through formal Lean proofs—machine-checkable code that either holds under scrutiny or does not. Each proof is reproducible by any mathematician, eliminating reliance on OpenAI's claims. The problems span territory that has occupied human experts for years; the $2,000 cost reframes advanced mathematics as something scalable with compute rather than constrained by scarce genius.

OpenAI deliberately chose to unveil Astra through mathematical discovery rather than benchmark scores, a deliberate departure from the usual model-launch playbook. This positions Astra as a scientific instrument for original research rather than a more capable conversational AI, signaling both the capability bar and the intended use case to Washington policymakers amid mounting AI oversight debates. The internal version is unreleased; public access timing remains unannounced.

For builders and research institutions, this signals a structural inflection: problems previously bounded by human researcher availability and months of effort now have a cost floor of $2K in compute. The implications ripple across academia, pharmaceutical research, materials science, and cryptography—any domain where formal mathematical proof is the bottleneck. Whether this generalizes beyond the areas where Astra excels remains the key question.