OpenAI's Astra Model Solves Ten Decade-Old Math Problems

AI Solves Ten Unsolved Mathematics Problems

OpenAI has announced solutions to ten mathematical problems that have remained unsolved for at least a decade—many for substantially longer. According to the OpenAI Blog, these results were generated by Astra, an internal unreleased model, and span high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics. The discoveries represent the first large-scale validation that large language models can contribute original proofs to fundamental open questions in mathematics.

Scope and Cost of the Breakthroughs

The problems addressed include sphere packing density optimization (advancing the Cohn–Elkies threshold), exponential improvements on binary and spherical code bounds, construction of non-sofic groups, a disproof of Connes’s rigidity conjecture, arithmetic circuit lower bounds, quantum parallel repetition theorems, hardness-of-approximation results for the closest vector problem (relevant to post-quantum cryptography), and progress on extremal combinatorics conjectures including Ehrhart’s volume conjecture and multicolor Ramsey numbers.

According to OpenAI’s announcement, the total compute expenditure to generate these solutions would cost roughly $2,000 at Sol API rates. This cost-to-breakthrough ratio—achieving results across eight distinct mathematical subfields with under $2,500 in inference compute—suggests that scaling language models for mathematical discovery may be economically tractable compared to traditional research funding models.

Formalization and Verification Process

A critical distinction from prior AI-assisted mathematics work is the integration of formal verification. After Astra generated mathematical arguments, human mathematicians prepared these results into peer-review-ready manuscripts. The model then formalized each argument in Lean, the interactive theorem prover, ensuring that proofs are machine-verifiable and resistant to subtle logical gaps. OpenAI is also releasing model narrations of its reasoning process for each solution, providing transparency into how the model approached the problems.

Institutional Support for AI-Assisted Mathematical Research

The announcements accompany OpenAI’s launch of ChatGPT for Academic Researchers, a program providing 100,000 scientists and mathematicians with free access to OpenAI’s best ChatGPT models. This initiative positions AI-assisted discovery as a core research tool, moving beyond individual researcher adoption to institutionalized access modeled on how computational tools became standard in mathematics and physics decades ago.

Why This Matters

These results validate the hypothesis that large language models can engage with deep mathematical structure, not merely retrieve or synthesize existing knowledge. For the mathematical community, the implication is immediate: Astra’s success on decade-old conjectures suggests that interactive theorem-proving workflows—where AI generates candidates and humans (or formal verification systems) validate them—may accelerate progress on the frontier of open problems.

For AI research, the breakthrough carries a secondary message: the cost-effectiveness of achieving mathematical breakthroughs ($2,000 per ten problems) suggests that inference compute, not training compute, is becoming the limiting factor for narrow, high-impact applications. Teams evaluating investments in theorem-proving infrastructure, post-quantum cryptography research, or pure mathematics will likely begin piloting internal versions of Astra or equivalent models, accelerating a shift toward AI-native mathematical discovery pipelines within the next 12–18 months.