OpenAI says an early, internal version of its next major model has done something no AI system has done before: produce new, verified solutions to ten mathematics and computer science problems that had gone unsolved for at least a decade. The company published the proofs alongside machine-checked certificates, letting anyone verify the logic step by step rather than take the claim on faith.
What OpenAI announced
On August 1, 2026, OpenAI said an internal version of its upcoming model family, code-named Astra, generated new results for ten open problems spanning pure mathematics and theoretical computer science (SiliconANGLE). According to the company, the model produced a 249-page collection of manuscripts, along with reasoning walkthroughs showing how it arrived at each result (TechTimes).
Every proof was formalized in Lean 4, a proof assistant that mathematicians use to check formal logic line by line. OpenAI reported a "sorry" count of zero across all ten submissions, meaning no step in any proof was left unproven or assumed — a detail that matters because it removes the possibility of hidden gaps that a purely narrative proof might hide.
The headline result
The most notable of the ten results is the first explicit construction of a non-sofic group. This resolves a question in group theory that has been open since 1999, when mathematician Mikhail Gromov introduced the concept of soficity (The Next Web). Sofic groups are a class of mathematical structures that has been difficult to fully characterize, and mathematicians have spent decades searching for a concrete example of a group that falls outside this class.
The other nine results touch areas including von Neumann algebras, sphere-packing density bounds, circuit complexity, and Ramsey theory — all active, technical corners of mathematics and computer science research.
Here's a quick breakdown of what was solved:
| # | Problem area | What it establishes |
|---|---|---|
| 1 | Group theory | First explicit construction of a non-sofic group (open since 1999) |
| 2 | Von Neumann algebras | New structural result in operator algebra theory |
| 3 | Sphere packing | New upper bound on sphere-packing density |
| 4 | Circuit complexity | New bound relevant to theoretical computer science |
| 5 | Ramsey theory | New combinatorial result |
| 6–10 | Mixed math/CS problems | Additional decade-plus-old problems, each with a zero-"sorry" Lean 4 certificate |
How much it cost
OpenAI put the total compute cost for generating all ten proofs at roughly $2,000, based on internal Sol API rates (DataCamp). That figure is likely to draw attention on its own: research-grade mathematical results that would typically take human specialists months or years to develop were produced, by OpenAI's account, for less than the price of a laptop.
| Detail | Figure |
|---|---|
| Problems solved | 10 |
| Oldest open problem resolved | 27 years (since 1999) |
| Manuscript length | 249 pages |
| Unproven steps ("sorry" count) | 0 |
| Total compute cost | ~$2,000 |
| Formal peer review completed | Not yet |
Why the verification matters
Claims about AI solving hard math problems are easy to overstate, which is exactly why the Lean formalization is the important part of this story. Lean proofs are checked by software, not by trusting a chatbot's explanation. If a Lean certificate compiles cleanly with no unproven steps, the logic is sound according to the rules of formal mathematics, independent of how the proof was generated.
That said, OpenAI's results have not yet gone through a formal, refereed journal review. Some mathematicians have reviewed preprints informally and found the work credible, but peer review is a separate and slower process that will determine how the broader mathematics community ultimately treats these results.
Context: what "Astra" is
Astra is described as OpenAI's next major model family, though the company has not detailed a full public release timeline. The math results come from an internal, pre-release build, so it is not yet clear how this capability will show up in any consumer or developer-facing product, or whether it reflects a narrow specialization rather than a general leap in reasoning ability.
The bigger picture
This announcement lands alongside a broader run of AI research news this week, including reports of coding agents modernizing old research software with significant speedups. Together, they point to AI labs increasingly using formal verification and measurable, checkable benchmarks — rather than open-ended demos — to back up claims about model capability. For a field that has faced criticism over inflated claims, machine-checked proofs are a harder standard to game, and that alone may be the more durable story here.
FAQ
What is Astra?
Astra is the code name OpenAI has given to its next major model family. The math proof results come from an internal, pre-release version of it.
What does it mean that the proofs are "Lean-verified"?
Lean is formal proof-checking software. A Lean-verified proof has been checked step by step by the software itself, not just written in natural language, which removes the risk of a subtle logical gap.
Has this work been peer reviewed?
Not yet through a formal journal process. Some mathematicians have reviewed preprints informally, but formal peer review is a separate, ongoing step.
What is a non-sofic group, in simple terms?
It is a mathematical structure that falls outside a well-studied class called "sofic groups." Mathematicians had been searching for a concrete example since the concept was introduced in 1999, and Astra's construction is the first one on record.
How much did this cost OpenAI?
The company says all ten proofs together cost roughly $2,000 in compute.
