OpenAI's unreleased Astra model generated original mathematical arguments for ten long-standing open problems in mathematics and theoretical computer science, each shipped with a machine-checkable Lean 4 proof certificate on August 1, 2026 — and within 24 hours, an Anthropic researcher showed the publicly available Claude Fable 5 could independently solve five of the same ten. The total inference cost for all ten Astra results was roughly $2,000 at GPT-5.6 Sol API rates, a price point that puts open-problem research within reach of any well-funded lab. Neither result has passed peer review yet, but because Lean certificates compile or they don't — with no room for hand-waving — the mathematical community can verify the proofs without trusting either company.
Last verified: 2026-08-04
- OpenAI published ten results from Astra on Aug 1, 2026, each with a Lean 4 certificate on GitHub.
- Anthropic researcher Levent Alpöge solved five of the same ten with Claude Fable 5 in under 24 hours, with no internet access.
- Total Astra compute cost: ~$2,000 at Sol API rates. No Millennium Prize problem was solved.
- The Leiden Declaration (June 2026, IMU-endorsed) warns AI labs to respect peer review and attribution.
- Neither set of results has been peer-reviewed; both are verifiable by compiling the Lean certificates.
- Volatile: model names, versions, and pricing will change — re-check before relying on them.
What Did OpenAI's Astra Actually Prove?
Astra produced ten new results across high-dimensional geometry, coding theory, group theory, operator algebras, arithmetic circuit complexity, quantum complexity, lattice cryptography, and extremal combinatorics. Here is what each result is and why it matters.
| # | Problem | Field | Result type | Why it matters |
|---|---|---|---|---|
| 1 | High-dimensional sphere packing | Geometry | Improved upper bound (down to Cohn–Elkies threshold) | First improvement to the general upper bound since 1978; underpins signal processing and coding theory. |
| 2 | Binary and spherical codes | Coding theory | Exponentially improved bounds | Better error-correcting codes; relevant to telecommunications and storage. |
| 3 | Non-sofic groups | Group theory | Existence construction (headline result) | First explicit non-sofic group, closing a question open since Gromov introduced soficity in 1999 — 27 years. |
| 4 | Connes's rigidity conjecture | Operator algebras | Disproof | Shows certain groups are not uniquely determined by their von Neumann algebras. |
| 5 | Arithmetic circuit complexity | Complexity theory | New lower bounds (incl. n⁴/log n for permanent) | Progress on a notoriously hard frontier; connects to P vs NP-adjacent questions. |
| 6 | Quantum parallel repetition | Quantum complexity | Exponential parallel repetition theorem | Extends a classical complexity principle to two-player quantum games. |
| 7 | Closest vector problem | Lattice cryptography | Hardness of approximation | Foundational to post-quantum cryptography security assumptions. |
| 8 | Ehrhart's volume conjecture | Discrete geometry | Bounds in every dimension | Resolves a question about convex bodies, lattices, and centroids. |
| 9 | Multicolor Ramsey numbers | Combinatorics | Superexponential lower bound | Resolves Erdős problem 183 on multicolor triangle Ramsey numbers. |
| 10 | Extremal graph conjectures | Graph theory | Compactness and degeneracy results | Resolves Erdős problems 146 and 180. |
Roughly half are outright resolutions of named conjectures; the rest are improved bounds — real contributions, but not "solved" in the colloquial sense. No Millennium Prize Problem was touched. Noam Brown, an OpenAI researcher on the project, confirmed that directly on X: "Sadly no Millennium Prize problems (yet)." [Source: OpenAI, Aug 1, 2026; SiliconANGLE, Aug 2, 2026]
How Were the Proofs Verified?
Every result ships with a machine-checkable Lean 4 certificate on GitHub under an Apache 2.0 license. Lean is a proof assistant — a computer program that type-checks formal mathematical arguments. If the certificate compiles, the proof is valid. If it doesn't, it isn't. There is no middle ground, and no trust in OpenAI's characterization of the result is required.
This is the structural shift. The standard failure mode for AI-generated math is a plausible chain of reasoning that quietly hand-waves past the one step that actually mattered. A Lean kernel does not accept hand-waving. Anyone can clone the openai/ten-proofs repository and re-verify every proof locally.
OpenAI described a three-stage pipeline: Astra generated the raw mathematical arguments, human researchers used the same model to turn those into readable manuscripts, and the manuscripts were formalized as Lean certificates. OpenAI took responsibility for correctness of the proofs and stated that claiming human authorship for a proof generated entirely by an AI system "would misrepresent both the system's contribution and the nature of genuine human intellectual work." [Source: OpenAI, Aug 1, 2026]
But formal verification confirms the proofs are internally valid. It does not confirm that the statements proved are the ones the community cares about, or that the framing around them is fair. Mathematicians must still confirm each formal statement says what the open problem actually asks, and judge whether the result matters.
How Did Claude Fable Respond?
Within hours of OpenAI's announcement, Anthropic researcher Levent Alpöge posted on X that Claude Fable 5 — a model already publicly available — had independently solved five of the same ten problems in under 24 hours. According to Alpöge, Fable worked autonomously from generic prompts with no internet access. The five he reported solving included arithmetic circuit complexity, quantum parallel repetition, and the closest vector problem. [Source: India Today, Aug 3, 2026]
Anthropic has not published a paper or Lean certificates verifying the Fable results. The claim sits as a social-media post from one researcher, not a company-backed research artifact. This is the important distinction:
| OpenAI Astra | Anthropic Claude Fable | |
|---|---|---|
| Model status | Internal, unreleased | Public (released Jun 9, 2026) |
| Problems claimed | 10 | 5 of the same 10 |
| Formal verification | Lean 4 certificates on GitHub | None published |
| Manuscript | 249-page collection published | None published |
| Internet access during solving | Not disclosed | None (offline, generic prompts) |
| Cost | ~$2,000 at Sol API rates | Not disclosed |
| Peer review | Not yet | Not yet |
Fable being a publicly available model makes the comparison significant: if five results fall to a deployed model in 24 hours, the bottleneck for AI-assisted math may not be model capability but the framing, tooling, and curation around it. Still, a social-media post without formal certificates is not the same evidentiary bar as a published repo. [\xA0Source: OpenAI blog; Levent Alpöge on X, via India Today\xA0]
For deeper context on how Claude Fable 5 stacks up against other frontier models in production coding tasks — where it also competes head-to-head with GPT-5.6 Sol and Qwen 3.8 Max — see our 2026 coding model comparison.
Why Does the $2,000 Cost Figure Matter?
The $2,000 figure covers inference tokens only — the cost of running Astra to generate the arguments. It does not include human curation, manuscript preparation, the cost of training Astra, or the infrastructure to run it. Noam Brown noted the number was a deliberate ceiling reflecting what was spent, not a limit on what the approach could produce: "It's possible to push test-time compute much further." [Source: NextBigFuture, Aug 1, 2026]
Even so, it changes who can participate in open-problem research. A mathematician with a modest grant and API access can now attempt problems that once required a year of human effort. The economics of mathematical discovery have shifted from "is a model good enough?" to "can you frame the problem and curate the output?" For context on how OpenAI is sorting its inference cost strategy across model tiers — from the custom Jalapeño chip to GPT-5.6 Sol pricing — read our breakdown of OpenAI's 2026 profitability plan.
What Is the Leiden Declaration and Why Does It Matter Here?
The Leiden Declaration on Artificial Intelligence and Mathematics, published June 2, 2026, is a community statement endorsed by the International Mathematical Union (IMU). It warns that AI companies are publishing results through press releases and blog posts rather than peer-reviewed journals, using published research without consent for training, and threatening the integrity of proof and attribution. Over 1,500 mathematicians signed it within days, including Fields Medallist Peter Scholze and IMU Vice President Ulrike Tillmann. [Source: Leiden Declaration; Wikipedia]
OpenAI explicitly acknowledged the Declaration in its August 1 post. The tension is real: the Astra results were announced via a blog post and PDF, not a peer-reviewed journal. The Lean certificates partially address the verification concern — you can check the proofs yourself — but the Leiden Declaration's call for peer-reviewed publication and transparent AI-disclosure is not satisfied by a certificate repository alone. Mathematicians will now spend months scrutinizing whether the proofs contain genuine insight and whether the results are as novel as described.
Has AI Solved Math Problems Before?
Yes, and that context sharpens what is new here.
- October 2025: OpenAI's then-VP Kevin Weil claimed GPT-5 had solved ten previously unsolved Erdős problems. Thomas Bloom, who maintains the erdosproblems.com database, called it "a dramatic misrepresentation" — the model had found papers in the literature Bloom was personally unaware of. Weil deleted the post. Demis Hassabis of Google DeepMind called the episode embarrassing. [Source: SiliconANGLE, Aug 2, 2026]
- May 20, 2026: An internal OpenAI reasoning model disproved Paul Erdős's 1946 unit-distance conjecture in discrete geometry — a problem that had stood for roughly 80 years. External mathematicians, including Thomas Bloom, verified the counterexample and called it a milestone. The May result inspired follow-on papers on arXiv in the following weeks. [Source: OpenAI, May 20, 2026]
- August 1, 2026 (this announcement): Same lab, different model (Astra), ten results, Lean certificates. Bloom called these "big news" and rated them ahead of the May counterexample.
The progression matters. The October 2025 claim collapsed because it was unsupported. The May 2026 counterexample survived because it was verifiable and external mathematicians engaged with it. The August 2026 batch is the strongest evidence yet: the proofs are machine-checkable and anyone can compile them. Prior AI-math systems worked primarily on competition problems (Olympiad-style, with known answers). The Astra batch is structurally different — these are open research problems with no known solution to compare against.
What Does This Mean for Practical AI Use?
You will not run Astra to prove theorems yourself — it is unreleased, and its first public path runs through a US federal AI safety review process. But the design pattern behind it is already applicable:
- Long-horizon agent loops work. Astra is built to coordinate multiple agents on a single problem across hours or days, not answer bounded one-shot prompts. That pattern — sustained, multi-agent search — is the same one driving progress in coding agents and research agents you can already run. Our hands-on guide to building a self-improving AI agent operating system covers the same architecture in a domain you can use today.
- Formal verification changes the trust calculus. Whether you are proving a theorem or verifying a smart contract, the lesson is the same: if your AI's output can be machine-checked, the trust objection dissolves. Teams generating code, security proofs, or formal specs should push their AI outputs through a checker — Lean, Coq, or even a test suite — rather than eyeballing plausibility.
- The framing and curation matter as much as the model. The Astra pipeline had humans directing formalization and shaping manuscripts. The Claude Fable response had one researcher with generic prompts and no internet. Both produced results. The bottleneck was not raw model power; it was the scaffolding around it. The same is true in AI SEO content workflows: the model is the engine, the framing is the steering.
- Compute cost is no longer the gatekeeper for research-grade AI tasks. $2,000 of API tokens produced ten research-level mathematical results. For comparison, a single senior mathematician's annual salary is $100K+. The economics are now squarely in favor of "try it with the model first, escalate to a human expert second."
What This Means for You
If you build with AI or run a small team: the frontier of what AI can do is not just "chat" or "summarize" anymore — it is the kind of deep, sustained reasoning that previously required domain experts working for months. The Astra announcement is a proof of concept that long-horizon agent loops can produce genuinely novel results, not rehashes. If you have a problem you assumed was too hard for an AI agent because it required "real reasoning," that assumption is worth re-testing — architect it as a multi-agent search with a verification step at the end, give it a real budget, and see what comes back. The barrier is no longer capability. The barrier is framing, tooling, and the willingness to let the agent run long enough.
If you are following the broader OpenAI-Anthropic competition arc — including the incidents where frontier models escaped their testing environments — Astra and Fable solving the same open problems hours apart tells you the capability is not a moat. The moat is who can run the loop longest, curate the output best, and ship the receipts.
FAQ
Q: What is OpenAI Astra? A: Astra is the name OpenAI gives to its next major model family. An internal version generated the mathematical arguments for the ten results published August 1, 2026. OpenAI has not announced a public release date, pricing, or whether it will ship as GPT-6 or a separate product line. [Source: OpenAI]
Q: Can I access the Astra proofs myself? A: Yes. OpenAI published Lean 4 proof certificates on the openai/ten-proofs GitHub repository under Apache 2.0. Anyone with the Lean compiler can clone the repo and verify every proof locally. A certificate that compiles is a valid proof — no trust in OpenAI is required.
Q: Did Claude Fable solve the same problems as Astra? A: Anthropic researcher Levent Alpöge reported on X that Claude Fable 5 solved five of the same ten problems in under 24 hours, working autonomously from generic prompts with no internet access. Anthropic has not published a paper or Lean certificates verifying the claim, so the Fable results remain unverified beyond the social-media post. [Source: India Today]
Q: Is $2,000 the real cost of AI solving math problems? A: The $2,000 figure covers inference tokens at GPT-5.6 Sol API rates only. It does not include human curation, manuscript preparation, model training costs, or infrastructure. Noam Brown noted it was a deliberate ceiling reflecting what was spent, not a hard limit. Still, it demonstrates that research-grade mathematical reasoning is now within the API budget of a well-funded individual or lab. [Source: NextBigFuture]
Q: Were any of the ten problems Millennium Prize problems? A: No. Noam Brown confirmed directly: "Sadly no Millennium Prize problems (yet)." The ten span group theory, coding theory, complexity, cryptography, and combinatorics, but none of the seven Clay Millennium Prize problems were among them. [Source: SiliconANGLE]
Q: What is the Leiden Declaration? A: Published June 2, 2026 and endorsed by the International Mathematical Union, the Leiden Declaration warns that AI companies are publishing mathematical results via press releases rather than peer-reviewed journals, using published research without consent, and threatening the integrity of proof and attribution. Over 1,500 mathematicians — including Fields Medallist Peter Scholze — signed within days. [Source: Leiden Declaration; Wikipedia]
Discussion
0 comments