OpenAI did not hold a press conference or run a launch campaign. On August 1, they quietly pushed a GitHub repository called ten-proofs containing Lean 4 formalizations of ten previously unsolved mathematics problems. The repository is signed by their next major model family: Astra.

This is not a benchmark result. These are real, machine-verifiable proofs of problems that had resisted the efforts of professional mathematicians for decades. The math community can check every step.

What Astra Actually Did

The ten results include:

• First-ever construction of a non-sofic group — a question open since Mikhail Gromov introduced the concept of soficity in 1999. For 27 years, no one could prove whether non-sofic groups even existed.
• Disproof of Connes’s Embedding Conjecture — related to fundamental questions in quantum mechanics and operator algebras.
• Resolution of three Erdős problems — including extremal graph theory conjectures.
• Sphere packing density bounds, quantum parallel repetition theorems, and several others.

Each result came with a 249-page paper and machine-checkable Lean 4 certificates hosted on GitHub. The compute cost to find all ten solutions was approximately $2,000 at OpenAI’s current Sol API rates.

This Is a Different Kind of Model

Astra is not GPT-5.x. It is a different architecture designed for long-horizon reasoning. Multiple AI agents coordinate over hours or days on the same problem. A root agent creates subagents, assigns work, waits for partial results, and synthesizes a final answer.

This is the same approach that is showing up across the most capable AI deployments right now — not a single model answering a question, but an orchestrated team of agents working a problem across time.

Sam Altman previewed Astra to US senators and White House officials on July 29, ahead of the formal August 1 deadline for the NSA to finalize AI pre-release review processes under the TRAINS executive order. Astra is expected to be an early test case for the new government evaluation framework.

What This Means for Business

Most business leaders are not going to deploy Astra to solve number theory problems. That is not the point.

The point is what solving decade-old mathematics problems at $2,000 of compute tells you about the direction of AI capability:

Long-horizon tasks are now tractable. The problems Astra solved required sustained reasoning over time, not just pattern matching. That same architecture — agents coordinating across a session lasting hours — is what makes complex business workflows automatable. Not simple tasks. Complex ones.

Verification matters more than generation. Every proof came with machine-checkable certificates. This is the pattern enterprises have been waiting for: AI that produces outputs that can be independently verified. In financial services, legal, healthcare, and audit — verifiability is the barrier to trust. Multi-agent systems with verification layers are how that barrier gets cleared.

The cost curve is brutal. $2,000 of compute to solve ten problems that stumped humanity’s best mathematicians for decades. Whatever you think about the capability, the economics are extraordinary. The question for every business leader is: what would have cost $2M of expert time two years ago, and what does it cost now?

What to Watch Next

Astra has not been released publicly. OpenAI is expected to submit it for federal review before launch — one of the first tests of the TRAINS pre-release evaluation process.

The math results are a signal. They demonstrate sustained multi-agent reasoning on hard, verifiable problems. The enterprise applications of that same capability — complex document analysis, multi-step research, long-horizon workflow automation — are not theoretical. They are already in deployment at the companies moving fastest on AI.

If your business is still thinking about AI as a tool for generating first drafts or answering FAQs, the gap between you and the leaders is widening faster than most forecasts predicted.

---

Enterprise DNA helps businesses build AI operations that match real capability, not last year’s hype. If you want to understand what current AI architectures actually mean for your workflows, book a workshop with Sam McKay.