Quick Answer
OpenAI’s unreleased Astra model solved 10 open math problems on August 1, 2026, each unsolved for at least a decade, producing machine-checkable Lean 4 proof certificates for roughly $2,000 in compute. The proofs are publicly verifiable on GitHub. Astra has no release date, no pricing, and must pass a US government security review before any public rollout.
Key Takeaways
- An internal version of Astra, OpenAI’s next major model, solved 10 long-standing problems spanning group theory, high-dimensional geometry, quantum complexity, and 5 other mathematical fields on August 1, 2026.
- Every proof is formalized in Lean 4 and published on GitHub under an Apache 2.0 license, so any researcher can verify them without trusting OpenAI’s claims directly.
- The headline result is the first-ever explicit construction of a non-sofic group, a question open since 1999 that no human mathematician had resolved in 27 years.
- OpenAI estimates the total token cost for all 10 solutions at roughly $2,000 at current GPT-5.6 Sol API rates, a figure that signals the compute requirement is within reach of any funded research lab.
- Astra has no confirmed public release date, no model card, and no pricing. It is expected to be the first model reviewed under the new US TRAINS government framework before launch.
What Did OpenAI’s Astra Actually Solve?
OpenAI’s Astra model produced 10 new results in mathematics and theoretical computer science, each addressing a problem that had seen no meaningful progress for at least a decade and, in several cases, for much longer. OpenAI published all 10 results on August 1, 2026, and each answered a problem that saw no progress on its main result for at least a decade. The announcement served a dual purpose: it demonstrated a significant capability advance and simultaneously gave the company’s next major model family its name.
The list covers a claimed counterexample to Connes’s rigidity conjecture on group von Neumann algebras, a construction establishing that non-sofic groups exist, improved asymptotic upper bounds for high-dimensional sphere packing down to the Cohn-Elkies threshold, exponentially stronger bounds for binary and spherical codes, an n⁴/log n formula lower bound for the permanent, and a superexponential lower bound for multicolor Ramsey numbers resolving Erdős problem 183. These are not incremental refinements of existing results. The OpenAI Astra math solutions are not benchmark scores. They are new mathematical results: constructions, counterexamples, and bounds that professional mathematicians had not been able to produce, now released with machine-checkable proofs in the Lean theorem prover.
OpenAI credited Astra with the underlying mathematical reasoning, while human researchers worked the model’s output into papers suitable for publication. OpenAI says the total tokens needed to find the solutions would cost roughly $2,000 at Sol API rates, putting the compute spend in a range any funded research lab could replicate. Humans then prepared the arguments into manuscripts using the same model, and the model formalized each one in Lean.
What Is the Non-Sofic Group Result, and Why Does It Stand Out?
The non-sofic group construction is widely considered the most significant of the 10 results. The headline result is the construction of the first known non-sofic group, a counterexample to a question that stood open since Mikhail Gromov introduced the concept of soficity in 1999. A group is sofic if its behavior can be approximated arbitrarily well by finite permutation systems; all amenable groups and all residually finite groups satisfy this condition. Whether every countable discrete group must be sofic was one of modern group theory’s most prominent unanswered questions for 27 years.
Astra’s construction answers it: no, not every group is sofic. The implications extend to operator algebra theory, ergodic theory, and the study of cellular automata. University of Manchester mathematician Thomas Bloom, who curates the erdosproblems.com database, weighed in publicly. Bloom called the results “big news” in an X post, adding that, as mathematical constructions, they were bigger than the unit-distance counterexample OpenAI had disclosed earlier.
The second most-discussed result concerns a foundational problem in algebra. The second major result is a disproof of the Connes Rigidity Conjecture, posed by Fields Medalist Alain Connes in 1980. The conjecture concerns whether the von Neumann algebra L(G) constructed from a property (T) group G retains enough information to reconstruct G. Astra disproved this conjecture by constructing infinitely many non-isomorphic groups with property (T) that share the same von Neumann algebra, and it also proved Ehrhart’s volume conjecture.
How Do the Lean 4 Proofs Change the Verification Picture?
The strongest reason this announcement differs from previous AI math claims is the Lean 4 formalization. These solutions are not incremental progress on the problems: they are genuine resolutions, verified with Lean. Lean is a programming language and proof assistant that forces every step of a mathematical argument to be spelled out in machine-readable detail. That requirement eliminates the main objection that has followed every previous AI-does-math claim.
Alongside the announcement, OpenAI released a 249-page manuscript and Lean 4 proof certificates on GitHub under an Apache 2.0 license; the repository’s “sorry” count stands at zero, indicating that every step across all ten formalized proofs is fully verified. In Lean, a “sorry” keyword is a placeholder that marks an unproven step; a count of zero means no logical gap remains in any of the 10 proofs. A Lean proof that type-checks is valid by construction. No trust in OpenAI’s evaluation claims, benchmarks, or interpretation is required. Any mathematician or independent researcher can pull the certificates from the openai/ten-proofs GitHub repository and verify them directly.
That shift in the verification structure matters for how readers should interpret the result. The relevant question is no longer whether the proofs are correct (they can be checked), but whether the mathematical ideas they contain are new and how they were found. The broader mathematical community is still reviewing the manuscripts and their implications, and full peer review has not yet been completed. That is a genuine limitation, and readers should distinguish between machine-verified formal correctness and full community acceptance of the underlying mathematical ideas.
What Is the $2,000 Compute Cost, and What Does It Mean for Research?
OpenAI put the total compute cost for all 10 solutions at roughly $2,000 at GPT-5.6 Sol API rates. That figure is OpenAI’s own estimate, stated as an approximation at a specific pricing tier as of August 2026, and actual costs would vary depending on the number of attempts, the problem setup, and how much human-guided iteration was involved. The $2,000 figure should be understood as a directional signal about accessibility rather than a precise reproduction cost.
The practical implication is meaningful nonetheless. A compute budget that any well-funded university research group could authorize is very different from the multi-million-dollar infrastructure that characterized earlier frontier AI research. This breakthrough highlights a shift where the cost of discovery is minimal, and the ability to guarantee correctness becomes paramount, enabling AI’s broader application in critical industries. At the same time, the cost estimate covers only the token usage to find the solutions. It does not account for the human expert time needed to select the problems, guide the model, and prepare the manuscripts.
OpenAI is also expanding researcher access beyond this announcement. The company is giving 100,000 academic researchers free access to its frontier models through 2027, deepening its ties to the scientific community while concentrating research infrastructure on its own platform. That program runs alongside the Astra math work but uses the currently available GPT-5.6 family, not the unreleased Astra model itself.
What Are the 10 Problems Astra Solved?
The table below summarizes each of the 10 results, the mathematical field it belongs to, and why researchers consider it significant. All descriptions are drawn from OpenAI’s August 1, 2026 publication and corroborating coverage from The Next Web, SiliconAngle, and The Decoder.
| Result | Field | Significance |
|---|---|---|
| First explicit non-sofic group construction | Group theory | Settled a question open since 1999; impacts operator algebras and ergodic theory |
| Disproof of Connes’s rigidity conjecture | Operator algebras | Shows certain groups are not uniquely determined by their von Neumann algebras; posed by Fields Medalist Alain Connes in 1980 |
| Proof of Ehrhart’s volume conjecture | Geometry | Resolves a longstanding conjecture in combinatorial geometry |
| Improved high-dimensional sphere-packing bounds (to Cohn-Elkies threshold) | High-dimensional geometry | First improvement to the general upper bound since 1978, per OpenAI’s manuscript |
| Exponentially improved bounds for binary and spherical codes | Coding theory | Advances maximum code-size bounds at any prescribed minimum distance |
| n⁴/log n lower bound for the permanent | Arithmetic circuit complexity | New lower bound on the circuit complexity of computing the permanent function |
| Superexponential lower bound for multicolor Ramsey numbers (Erdős problem 183) | Extremal combinatorics | Resolves a problem from Paul Erdős’s famous catalog |
| Results on compactness and degeneracy conjectures (Erdős problems 146 and 180) | Extremal combinatorics | Resolves two additional problems from the Erdős catalog |
| Parallel repetition theorem for two-player quantum games | Quantum complexity | Advances understanding of entanglement and complexity in quantum information theory |
| New lattice cryptography results | Lattice cryptography | Relevant to post-quantum cryptographic security foundations |
Sources for the above: The Decoder’s coverage of the August 1 announcement, The Next Web’s detailed breakdown, and SiliconAngle’s report.
What Are the Legitimate Criticisms and Limitations?
The Lean 4 formalization addresses the question of proof correctness, but several other concerns remain open. OpenAI disclosed almost nothing about the model, its training process, the number of failed attempts, or the total human effort involved. That gap creates the central conflict. The disclosed results show that AI can make meaningful contributions within selected, formal problems. The selection of problems is itself a point of scrutiny: OpenAI chose which problems to attempt, and researchers do not have visibility into how many attempts were made or how many problems were tried but not solved.
Critics question problem selection and model access, though the verifiable nature of Astra’s output addresses a key challenge in AI deployment. AI researcher Gary Marcus has called the results remarkable but cautioned that they are, in his view, “vastly oversold” as evidence of general reasoning ability. Machine checking is not the same as complete community validation. Full peer review of the underlying mathematical ideas and their originality has not yet occurred, and the broader mathematical community is still working through the 249-page manuscript collection.
One additional historical note is relevant context. In October 2025, an announcement in which OpenAI claimed GPT-5 had solved ten Erdős problems collapsed within days when those claims could not be independently verified. The August 2026 announcement is structurally different because the Lean 4 certificates allow direct machine verification, but the contrast with the prior retracted claim is worth keeping in mind when reading near-term coverage.
When Will Astra Be Available, and Will It Be GPT-6?
Astra is not available to the public as of August 6, 2026, and OpenAI has not announced a release date. Astra is not released, with no date, no pricing, no model card and no ChatGPT availability, and OpenAI has not said whether it ships as GPT-6 or as another GPT-5 point release. The public record does not yet include Astra’s release date, model sizes, context window, API price, ChatGPT plans, safety card, tool support, or whether the research system maps directly onto the product users will eventually receive.
The regulatory pathway adds a meaningful but unquantified delay. Astra is the first model family designated for the TRAINS pre-release national security evaluation process, established under Executive Order 14409. The review scope and timeline are not public, and how long it runs will shape when Astra reaches API access. Sam Altman gave US lawmakers and regulators a private Astra demo in Washington in late July 2026. That briefing suggests the review process was already underway before the public announcement, but no timeline has been confirmed.
On the GPT-6 question specifically, OpenAI has also not decided whether Astra would be labeled GPT-6 or an additional model in the GPT-5 series, such as GPT-5.7. The project reflects OpenAI’s broader ambition to build AI systems capable of working on problems continuously for hours or even days at a time. That architectural description, multiple agents coordinating over extended periods rather than answering single prompts, is distinct from the current GPT-5.6 Sol, Terra, and Luna family that is available today. For users following OpenAI’s model roadmap, the most accurate summary is that Astra is named, internally tested, and mathematically notable, but it is not a product you can access. See our earlier coverage of OpenAI’s GPT-5.6 Luna pricing changes for context on the currently available model tier.
What Does This Mean for Everyday AI Users and Researchers?
For most general users of ChatGPT or the GPT-5.6 API today, the Astra math results do not change anything about the tools available. The model that produced those results is not accessible, and no release timeline exists. The practical significance is forward-looking: it signals the direction OpenAI is pursuing for its next generation of models, particularly in long-horizon reasoning tasks that require sustained work across multiple steps rather than a single prompt-and-response exchange.
For researchers, the most immediately actionable element of the announcement is the publicly available proof repository. OpenAI’s public repository, created at 06:10 UTC on August 1, 2026, uses Lean 4.32.0 with the mathlib library and Lake build system. Its README gives two commands for downloading cached dependencies and building all certificates. The Apache-2.0 license allows others to inspect and run those files without access to Astra. Any mathematician or developer with Lean 4 installed can verify the proofs independently today, regardless of when or whether Astra becomes publicly accessible.
The broader pattern for readers following AI development is also worth noting. OpenAI has described a research roadmap that targets research-intern-level AI capability as early as September 2026 and a fully autonomous AI researcher by early 2028. The Astra math results are the strongest public evidence so far supporting the near-term portion of that timeline, but the timeline remains OpenAI’s stated goal rather than a confirmed milestone. For related context on how AI tools are entering research and professional workflows, see our coverage of Google’s Gemini AI in K-12 classrooms and ChatGPT’s expanded financial data integrations.
FAQ
Are the Astra math proofs really verified, or is this just another AI math claim?
The proofs are verified in the strict Lean 4 sense: OpenAI released Lean 4 proof certificates on GitHub under an Apache 2.0 license, and the repository’s “sorry” count stands at zero, indicating that every step across all ten formalized proofs is fully verified. That machine-level verification is independent of OpenAI’s own claims. The remaining open question is whether the broader mathematical community will accept the underlying ideas as genuinely novel after full peer review, which is still ongoing.
Can I use Astra right now?
Astra is not publicly available as of August 6, 2026. OpenAI has not announced a release date, a model card, pricing, or ChatGPT availability for Astra, and the company has not confirmed whether it ships under the GPT-6 label or as a GPT-5 variant. The Lean 4 proof files on GitHub are publicly accessible under an Apache 2.0 license, but the model itself is not.
Does Astra solving these problems mean AI has achieved AGI?
Solving 10 formal mathematical problems does not constitute artificial general intelligence by any standard definition. Solving problems that resisted mathematicians is a major capability signal. It still does not meet the classic broad definition of superintelligence. The useful concept in between is jagged, domain-superhuman intelligence. Astra performed at an exceptional level on a specific class of formal, machine-verifiable mathematical tasks; its performance across other domains remains undisclosed.
What is the TRAINS review, and why does it matter for Astra’s release?
Astra is the first model family designated for the TRAINS pre-release national security evaluation process, established under Executive Order 14409. The review scope and timeline are not public, and how long it runs will shape when Astra reaches API access. This process represents a new gate that did not apply to earlier GPT-5 releases, and it could meaningfully affect how long it takes before Astra becomes available to developers or general users.
What is Lean 4, and why does it matter for these proofs?
These solutions are genuine resolutions verified with Lean. Lean is a programming language and proof assistant that forces every step of a mathematical argument to be spelled out in machine-readable detail. That means no step can be skipped or hand-waved. Previous AI math outputs that looked like proofs but could not be independently checked by machine left room for the objection that the model was pattern-matching rather than reasoning. Lean 4 eliminates that objection for these results.