OpenAI Astra Solves 10 Open Math Problems for $2,000

OpenAI Astra Solves Ten Decades-Old Open Problems in Mathematics

On 1 August 2026, OpenAI revealed that an internal version of Astra — the company's next major model, not yet available as a public API — had produced solutions to ten open problems in mathematics and theoretical computer science, each of which had resisted progress for at least a decade and in most cases far longer. OpenAI published the results as a 249-page manuscript alongside machine-checkable Lean 4 proof certificates on GitHub under an Apache 2.0 licence. The repository's "sorry" count stands at zero, which in Lean 4 means that every step of every formalised proof has been independently verified by the proof assistant with no gaps left open. Greg Brockman noted that the total compute required to reach all ten results would cost roughly 2,000 US dollars at current Sol API rates.

Six Fields, Ten Results

The ten results span six fields: group theory, operator algebras, high-dimensional geometry, quantum complexity, extremal combinatorics, and circuit complexity. The most prominent result is the first explicit construction of a non-sofic group — a proof that at least one group exists which cannot be approximated by symmetric groups in the sense Mikhail Gromov defined when he introduced the concept of soficity in 1999. The soficity question had been a central open problem in group theory for the 27 years since Gromov's original paper, and Astra's construction settles it.

A second major result is a disproof of Connes's rigidity conjecture. Astra constructed infinitely many non-isomorphic groups with property (T) that share the same von Neumann algebra — a result that contradicts the conjecture that the group structure of property (T) groups is rigidly determined by their von Neumann algebra invariants. Several of the ten results also resolve problems posed by Paul Erdős across the combinatorics and geometry domains.

Lean 4 Certificates and Independent Verifiability

The decision to publish formal Lean 4 certificates for every result is technically significant. A Lean 4 certificate is machine-checkable: any researcher with the Lean 4 proof assistant installed can verify the entire argument from axioms to conclusion without relying on human peer review. The zero "sorry" count signals that no step in any of the ten proofs has been left as an admitted gap. Fields Medal winner Timothy Gowers said he would recommend one of the Astra proofs for publication in a top mathematics journal without hesitation.

What Astra Is and When It Will Be Available

OpenAI has not published a release date for Astra and the model is not available through any public API or consumer product. The 1 August announcement framed the mathematics results as a first public disclosure of Astra's existence, demonstrating its capacity to reason through extended, multi-step, formally rigorous arguments over open research questions — a category of task qualitatively distinct from assisting with known proofs or symbolic manipulation using current frontier models.

The Cost Argument: 2,000 US Dollars for Decade-Old Breakthroughs

Greg Brockman's statement that all ten results cost approximately 2,000 US dollars in compute at Sol API rates is a deliberate contrast with the resources traditionally required at this level of mathematical research. Research careers are measured in years, conference budgets in hundreds of thousands, and grants in millions — for work that may or may not reach the frontier. A 2,000 US dollar compute budget delivering ten machine-verified open-problem resolutions across six fields is a data point about the changing economics of mathematical discovery, not a claim that human expertise is redundant.

What This Means for Indian Software and AI Product Teams

For Indian AI product companies and software teams, the Astra announcement matters on two levels. At the capability level, it signals that the next OpenAI model tier will handle extended autonomous reasoning tasks well beyond what GPT-5.6 Sol currently manages — with direct implications for AI-assisted legal document analysis, financial modelling, scientific research workflows, and complex software design tasks that Indian teams are building for enterprise clients. At the cost level, Brockman's 2,000 US dollar figure continues the pricing compression trend that has made frontier AI inference accessible to companies that are not hyperscalers: each successive capability jump is arriving at a lower compute cost than the one before it.

The Bottom Line

On 1 August 2026, OpenAI publicly disclosed Astra — its next major model — by publishing solutions to ten decades-old open problems across six fields: group theory, operator algebras, high-dimensional geometry, quantum complexity, extremal combinatorics, and circuit complexity. The headline results include the first explicit non-sofic group construction since Gromov's 1999 paper and a disproof of Connes's rigidity conjecture. All ten results are published as a 249-page manuscript and machine-verifiable Lean 4 proof certificates on GitHub under Apache 2.0, with a "sorry" count of zero. Fields Medal winner Timothy Gowers endorsed one result for top journal publication. Greg Brockman said the full set required compute costing approximately 2,000 US dollars at Sol API rates. Astra has not yet been released publicly.