OpenAI's Astra: 10 Frontier Math Proofs for ~$2,000 — and Anthropic's 24-Hour Reproduction

> Tip: Core keywords — OpenAI Astra, Lean 4 formal verification, Kun-Thom expanders, Thompson's group V, Connes rigidity conjecture, Cohn-Elkies linear programming, Erdős 183/146/180, Anthropic Fable, $2,000 compute bill.

The $2,000 Compute Bill and the Math Community's Priority Anxiety

On August 1, OpenAI researcher Sébastien Bubeck announced on X that the unreleased next-generation model Astra had proven ten frontier mathematical results, each open for at least ten years — accompanied by ten Lean 4 certificates, ten per-problem solution walkthroughs, and a 249-page technical manuscript. The figure everyone fixated on: $2,000, OpenAI's back-of-envelope token cost for the ten successful solution paths at GPT-5.6 Sol API pricing.

The details that actually stung were not mathematicians but engineers and PMs who treated reasoning models as fancy autocomplete. Someone pointed the models at Grothendieck-style pure structure problems and fed every step to Lean 4 for independent machine verification.

> Tip · Lean 4: A formal proof language. Mathematical statements and inference steps are written in machine-checkable syntax; every line is either accepted or rejected by the rules. For three years, AI math was mocked as "pretty conclusions with a skipped step" — Lean 4 certificates are nearly the only ticket back into serious conversation.

But the $2,000 figure needs an asterisk. Developer Simon Willison immediately noted that OpenAI published the *token cost of successful solutions*, without saying how many attempts failed at similar cost, or accounting for training compute, researcher effort in problem selection and write-up, or the engineering behind the Lean libraries. "Cost of the success path" ≠ "cost of frontier math."

The Ten Problems: Group Theory, Geometry, Cryptography, Combinatorics

| # | Problem | Field | Age | Core method | |---|---|---|---|---| | 1 | High-dimensional sphere packing | Analytic number theory + geometry | 1978–2026 (46 yrs) | Cohn-Elkies LP + exact exponential decay rate | | 2 | Binary and spherical codes | Coding theory | 1970s– | Exponential improvement on classical bounds | | 3 | Existence of non-sofic groups | Group theory / operator algebras | 1999–2026 (27 yrs) | Binary Leavitt algebra unit groups + Property (T) expanders + Thompson's group V | | 4 | Connes rigidity conjecture | Operator algebras | 1980s– | Infinitely many non-isomorphic groups sharing one von Neumann algebra | | 5 | Arithmetic circuit complexity | Theoretical CS | 1960s– | New circuit lower bounds for the permanent | | 6 | Quantum parallel repetition | Quantum information | Extension of a classic | Finite-field extension of two-player entangled games | | 7 | Closest Vector Problem (CVP) | Lattice cryptography | 2000s– | Polynomial-factor hardness + cryptographic consequences | | 8 | Ehrhart volume conjecture | Ehrhart theory | 2000s– | Sharp volume bounds in arbitrary dimension | | 9 | Multicolor Ramsey numbers | Combinatorics | Erdős problem 183 | Super-exponential lower bounds | | 10 | Extremal graph degeneracy | Extremal graph theory | Erdős problems 146, 180 | Counterexample construction |

> Tip · Sofic groups: An abstract property proposed by Gromov in 1999 — a group is sofic if it can be approximated by finite symmetric groups. It underpins sofic entropy, ergodic theory, and operator algebras; if a non-sofic group exists, the theoretical framework must be rewritten.

Problem #1 made headlines because it touches Fields-Medal-level density questions. Viazovska solved dimensions 8 and 24 (Fields Medal, 2022), but the 1978 Soviet upper bound as dimension → ∞ had not been improved by a single digit in 46 years. Astra not only gave a new proof but precisely computed the exponential decay rate of the Cohn-Elkies linear program.

Problem #3 is in Cambridge Fellow Henry Bradford's own field; he described the breakthrough in press pieces as something he "would have found unbelievable a month ago." Problem #4 disproves a famous rigidity conjecture by 1982 Fields Medalist Alain Connes. Problem #7 (CVP polynomial-factor hardness) pulls the story back to applications: lattice cryptography's security rests on CVP being hard.

Problem #3: An Engineering-Style Breakthrough

Astra's non-sofic group construction chained four steps:

1. Start from the unit group of the binary Leavitt algebra (Leavitt, 1960s). 2. Embed it into Property (T) expander graphs — Margulis's rigidity property is nearly the opposite direction from soficity, so "compressing" the algebraic structure into graph structure is the key translation. 3. Force in Thompson's group V, the strange self-similar infinite group studied by Richard Thompson in the 1960s, embedding its fine automorphisms into irreducible representations of the Leavitt unit group — producing a hybrid extreme in both directions. 4. Translate everything into Lean 4. The ten certificates contain zero "sorry" placeholders — no skipped steps anywhere.

This is not a new mathematical concept but a combination of existing tools in a form nobody had imagined.

24 Hours Later: Anthropic's Fable Reproduces Half

On August 2, less than 24 hours after the announcement, Anthropic researcher Levent Alpöge replied under Bubeck's post: "I did half of them with Fable."

Fable is not an unreleased model — it is publicly available. Alpöge described clean experimental conditions: fully autonomous, generic prompts, no web access, with safeguards preventing OpenAI's solutions from leaking into context. He reproduced items #4–#8: the Connes rigidity counterexample, arithmetic circuit lower bounds, quantum parallel repetition, CVP hardness, and the Ehrhart volume conjecture — 5 of 10.

The structural significance: the shelf life of a "first" has shrunk from years to days.

• Priority competition shifts from outcome-based to speed-based: the winner is not who publishes first but who is independently verified first.
• "I can use it" matters more than "I can build it." OpenAI's unpublished Astra held its first-mover lead for only 24 hours; the model that caught up was publicly available to everyone.

The Hidden Other Half of the $2,000 Ledger

Hidden cost 1 — training compute. Pretraining and post-training stages cost millions to tens of millions of dollars even for smaller frontier models; this was absorbed into OpenAI's research budget, not the "solving cost."

Hidden cost 2 — the failure ledger. Ten successful paths ≠ ten problems' total attempts. Typical research loops may need ~100 attempts per success, each consuming tokens, Lean compile time, and researcher filtering time. OpenAI published walkthroughs but not prompts, so the number of attempts per problem cannot be independently verified.

Hidden cost 3 — researcher brainpower. OpenAI acknowledged the arguments came from Astra but the transformation into publishable manuscripts was human work — not folded into the $2,000.

> Tip · Research accounting: Traditional funders (NSF, ERC, JSPS) evaluate "opportunity cost + failure cost + coordination cost" in total. Isolating the success path is a demo logic, not a scientific evaluation logic.

The Identity Crisis: From Solving Machine to Discovery Partner

On August 27, major outlets published mathematicians' responses. Henry Bradford (Cambridge) wrote:

> "Like Kasra Rafi and Bruce Schneier, I think recent AI breakthroughs in math are essentially clever recombination of existing theory rather than genuinely new theoretical development. But a month ago I would not have believed AI could do this — so betting against superhuman AI capability across all areas of mathematical thinking in the next few years now seems unwise."

His argument challenges the community's foundations: if AI produces theorems faster and cheaper, what is the *raison d'être* of mathematical research? Bradford invokes Bill Thurston's 1994 essay *On Proof and Progress in Mathematics*: mathematicians want not a collection of answers but ideas living in human minds — theorems are only markers for sharing ideas.

Two futures:

• Path A — solving machine + acceptor: AI produces theorems, mathematicians audit proofs, universities pay for "output," the community shrinks.
• Path B — discovery partner + idea source: AI explores and constructs, mathematicians choose problems and supply intuition, universities pay for ideas, the community expands.

A Cambridge/HKU-led arXiv paper (2608.23691, *Autonomous Mathematical Discovery in Open-World Multi-Agent Environments*) reported that research agents given free choice of exploration direction produce proofs and generalizations their evaluators never asked for — gains come mainly from spending compute on *structural exploration* rather than *search*, consistent with Bradford's "discovery partner" framing.

Formal Verification ≠ Peer Review: What Lean 4 Solves and Doesn't

What Lean 4 solves: whether each step of the reasoning chain is accepted by the rules — no skipped logic, no unproved lemmas, no phantom citations. Zero "sorry"s means no placeholder steps.

What it doesn't solve:

• The *mathematical meaning* of the statement — Lean only validates "if premises hold, conclusion follows."
• *Translation fidelity* — formalizing Erdős problem 183 is human work; a mistranslation yields a perfect certificate for the wrong claim.
• *Importance judgment* — purely academic, beyond Lean's reach.

Thomas Bloom (University of Manchester, Erdős problems database maintainer) called the results "big news" but stressed: "I haven't seen independent review, so these are *reported*, not *confirmed*, results."

More precisely: Astra's ten results are the first batch of outcomes that are machine-verified as logically airtight but still await human mathematical review. The review bottleneck shifts from "does the proof have holes" to "does the community care about the question."

Watchlines for the Next 6–12 Months

1. Weekly-cadence math AI releases — journals may need fast tracks for machine proofs; mathematicians' time shifts toward evaluating questions rather than proving them. 2. Formalization library coverage — Astra could solve non-sofic groups because Lean libraries for Leavitt algebras, Property (T), and expanders were mature. Coverage bounds AI math capability. 3. Whether Anthropic / xAI / DeepMind adopt the "announce + open reproduction" dual-track strategy — Fable's 5/10 in 24 hours shows public models with generic prompts already reach this frontier. 4. University budget battles — whether UK/EU/US universities cut pure-math positions on the assumption that mathematicians are replaceable. 5. An "idea economy" — markets for valuable unproven conjectures, benchmarks for mathematical intuition.

Conclusion: Two Doors for the Math Community

The week of August 27 was a watershed. OpenAI's Astra cracked ten hard problems at once; Anthropic's Fable reproduced half within 24 hours. "AI attacking frontier pure math" is no longer an occasional lab event — it is a publicly verifiable capability.

Two doors now stand before mathematicians:

• Become acceptors of AI proofs — easy, since review is reduced to judging whether questions matter.
• Become idea sources for AI — hard, because universities don't pay for ideas and AI can't yet encode intuition into testable claims.

Will the $2,000 ledger become the occasion for rewriting review standards and defending research freedom — or the last straw for pure-math positions? That is the question to track over the next 6–12 months.

References

1. OpenAI, *Ten new proofs, ten new problems* — https://cdn.openai.com/pdf/ten-proofs-oai.pdf (249-page technical manuscript) 2. OpenAI public repository — https://github.com/openai/ten-proofs (Apache 2.0, Lean 4 formal certificates) 3. Fello AI, *OpenAI Astra: What It Is and Why It Is Paused* — https://felloai.com/openai-astra 4. Inside AI News, *AI Proof of Non-Sofic Groups Forces Mathematicians to Rethink Purpose* (2026-08-27, Henry Bradford) 5. World Programming, *Surprising AI breakthroughs raise soul-searching questions for mathematicians* 6. Morning Overview, *OpenAI's Astra AI cracked 10 unsolved math problems for just $2,000 in computing* (2026-08-27) 7. Sohu, coverage of Astra's 10 proofs and Fable's 24-hour reproduction of 5 (Levent Alpöge reproduction record) 8. forRide, analysis of the 249-page PDF (Thompson's group V and Kun-Thom expanders) 9. Wortins, *OpenAI's Astra solves 10 long-unsolved math problems, publishes verified proofs* (2026-08-27) 10. arXiv 2608.23691, *Autonomous Mathematical Discovery in Open-World Multi-Agent Environments* (Cambridge/HKU-led labs, 2026-08-27)