[Deep Dive] OpenAI's Astra Solved 10 Unsolved Math Problems for $2,000 in Compute
![[Deep Dive] OpenAI's Astra Solved 10 Unsolved Math Problems for $2,000 in Compute](/content/images/size/w1200/2026/08/deep_dive_thumb-2.png)
OpenAI's Astra Solved 10 Unsolved Math Problems for $2,000 in Compute
AI Research β’ August 02, 2026
Reading time: ~12 minutes
π Contents
π Executive Summary
On August 1, 2026, OpenAI researcher Noam Brown said an internal build of Astra, the company's next model family, had closed ten open problems across mathematics and theoretical computer science, each unsolved for a decade or longer. The headline item is an explicit construction of a non-sofic group, answering a question Mikhail Gromov raised in 1999. Astra also reported a refutation of the Connes rigidity conjecture, solutions to three Erdos problems, and progress on high-dimensional sphere packing. Total inference spend: roughly $2,000 at Sol API rates. The pipeline matters as much as the results. Astra produced the arguments, worked with human co-authors on a 249-page manuscript, then formalized each proof into a machine-checkable Lean 4 certificate published to GitHub. Astra is not a shipping product. What shipped is a cost datapoint: frontier research output priced in hundreds of dollars per theorem, with verification cheap enough to be non-negotiable.
Ten problems, each unsolved for at least a decade, closed for roughly $2,000 in compute. That works out to about $200 per theorem, which reprices not mathematics but the question of who can afford to attempt it.
π¬ Technical Deep Dive
Current State
Two years ago, AI mathematics meant competition problems. AlphaProof and AlphaGeometry hit silver-medal level at the 2024 International Mathematical Olympiad; OpenAI and Google DeepMind both reported gold-medal performance in July 2025 with general reasoning models rather than bespoke solvers. Those were closed-form problems with known answers, graded against a rubric. The interesting shift through late 2025 and the first half of 2026 was the move from graded problems to unsolved ones, and from natural-language plausibility to machine-checkable certificates.
That shift had a rough start. In October 2025, an OpenAI staffer claimed GPT-5 had solved ten Erdos problems, which turned out to be literature retrieval of existing solutions the erdosproblems.com database had marked open. The retraction was fast and public, and it set the bar for everything after: no formal artifact, no claim. The Astra announcement is engineered around that lesson. Every one of the ten results ships with a Lean 4 proof term, so the audit surface is the kernel and the axioms file, not the prose.
| Dimension | Frontier state, mid-2025 | Astra, reported August 2026 |
|---|---|---|
| Problem type | Olympiad and benchmark problems with known answers | Decade-plus open problems in research literature |
| Agent horizon | Single agent, minutes to a few hours | Coordinating agents planning, revising, delegating over hours to days |
| Output artifact | Natural-language proof sketch, human-reviewed | Manuscript plus per-theorem Lean 4 certificate on GitHub |
| Verification cost | Weeks to months of expert referee time | Kernel check in minutes; refereeing shifts to definitions and statements |
| Reported spend | Undisclosed; IMO runs widely estimated in the tens of thousands | ~$2,000 for ten results |
| Human role | Prompt, verify, rewrite | Co-author on exposition; adjudicate whether formal statement matches the real question |
Read the table as a change in where the bottleneck sits. It used to sit in generation. It now sits in specification.
Recent Breakthroughs
The non-sofic group is the result specialists care about. A group is sofic if its finite chunks can be approximated by permutations of a finite set, with errors that vanish as the set grows. Gromov floated the concept in 1999 while studying symbolic dynamics, and soficity quickly became load-bearing: Gottschalk's surjunctivity conjecture, Kaplansky's direct finiteness conjecture, and Lueck's determinant conjecture are all known for sofic groups. Nobody could produce a group outside the class. Every amenable group is sofic, every residually finite group is sofic, and the class is closed under most standard operations, so the search space for a counterexample was uncomfortably thin.
The nearest precedent is the 2020 construction by De Chiffre, Glebsky, Lubotzky and Thom of groups not approximable by unitary groups in Frobenius norm, which used a cohomological rigidity obstruction: take a lattice with strong property (T)-style vanishing, build a central extension, and show any near-representation must be perturbable to an exact one, which the extension forbids. Adapting that machinery from unitary approximation to permutation approximation was the standing open program, and the obstacle was that Hamming-metric stability behaves worse than Frobenius stability. From the manuscript's structure, Astra's construction runs in that lineage, combining a rigidity input with an explicit extension and a stability argument that closes the permutation case. The Connes rigidity refutation follows a similar flavor of reasoning in operator algebras: exhibiting non-isomorphic icc property (T) groups whose group von Neumann algebras coincide, cutting against the W*-superrigidity direction that Chifan, Ioana, Osin, Popa and Sun had been extending since 2021.
The other seven results are less glamorous and arguably more informative about capability. Progress on high-dimensional sphere packing, three Erdos problems, arithmetic circuit complexity, coding theory, lattice cryptography and quantum complexity suggests a system that generalizes across proof cultures rather than one tuned to a single subfield. The $2,000 figure is the part that reframes everything. If it holds under scrutiny, a research-grade theorem now costs about what a conference registration costs, and the constraint on mathematical output becomes the supply of well-posed questions.
Remaining Challenges
Lean 4 certificates verify that a proof follows from stated definitions. They do not verify that the definitions encode the question the community cares about. Formalization error moves upstream: a subtly weakened hypothesis, a non-standard definition of soficity, an axiom import that trivializes a step. The Equational Theories Project and the Lean mathlib community have both documented cases where a formal statement passed the kernel while drifting from intent. With a 249-page manuscript and ten independent formalizations, that audit is weeks of expert work, and only a few dozen people worldwide can competently check the operator algebras piece.
Second, reproducibility. Astra is unreleased. Nobody outside OpenAI can rerun the ten problems, vary the seed, or measure how many attempts failed. The $2,000 figure almost certainly prices successful runs at API rates rather than the full search, including dead ends, human steering time, and the amortized training cost of the model family. Treat it as a marginal-cost datapoint, not a total cost of ownership.
Third, selection. Ten problems out of an attempted set of unknown size, chosen and announced by the vendor, tells you about the ceiling and nothing about the hit rate. Open problems vary enormously in how much they reward a system with encyclopedic literature recall and high-throughput construction search versus one needing a genuinely new conceptual frame. Constructions and counterexamples, which is what most of these ten are, sit closer to the machine's strengths than a program like resolving the Langlands correspondence.
Expert Perspectives
Thomas Bloom, the University of Manchester mathematician who maintains the Erdos problems database, called the results 'big news' and a 'significant step' for the field of constructions. That framing is precise and worth holding onto: constructions and counterexamples are where automated search plus formal checking has structural advantage, because success is a witness that verifies itself.
Terence Tao has spent two years arguing for exactly this division of labor, describing current models as useful collaborators for lemma-scale work and formalization while remaining unreliable at strategic direction. His view through 2025 was that AI would first industrialize the mid-tier of mathematics, the thousands of technically hard but conceptually routine results that gate progress. Ten decades-old problems is a faster arrival than that timeline implied. Skeptics in the formal methods community, including voices around the Lean FRO, have been consistent that certificates are necessary and insufficient, and that the community should insist on independent statement review before the word 'solved' gets used. Both camps agree on the practical near-term consequence: refereeing is about to change more than research is.
π’ Market Landscape
Key Players
The autonomous-research race now has three tiers. Frontier labs treat mathematics as a proving ground for long-horizon agents, because proofs are the only knowledge work with a cheap oracle. Specialist startups sell formal reasoning as infrastructure. And the AI-for-science layer applies the same agent patterns to chemistry, materials and biology, where verification is a lab bench rather than a kernel.
| Player | Math/reasoning asset | Verification approach | Status |
|---|---|---|---|
| OpenAI | Astra family; multi-agent, day-scale horizons | Lean 4 certificates, GitHub-published | Internal, not released as a product |
| Google DeepMind | AlphaProof, AlphaEvolve, Gemini Deep Think | Lean-based proving plus evolutionary program search | Partial access via trusted-tester programs |
| Anthropic | Claude Opus/Sonnet reasoning line, Agent SDK | Tool-use and code execution; no dedicated prover line | Commercially shipping |
| Harmonic AI | Aristotle, formal-first prover | Lean-native, hallucination-free positioning | Public API and app |
| Axiom Math | Self-improving mathematician agent | Formal proof search | Early stage, seed funded |
| Math Inc. | Gauss autoformalization agent | Large-scale Lean formalization pipelines | Early commercial |
| Open source | DeepSeek-Prover line, Qwen and Kimi reasoning models | Lean 4 RL training on formal corpora | Weights available |
The strategic asymmetry: OpenAI and DeepMind can afford to run day-long agent swarms as a research expense, while startups must monetize the formal layer immediately. Harmonic raised roughly $100 million in mid-2025 at an $875 million valuation on exactly that thesis. Axiom Math, founded by Carina Hong, closed a $64 million seed in 2025. Math Inc. is commercializing autoformalization as a service, which is arguably the most defensible position on the board, because every Astra-style result creates demand for someone to check it.
Investment Trends
Money followed the agent thesis before this announcement, not because of it. Global AI startup funding ran above $200 billion in 2025 by most trackers, with agentic and research-automation companies capturing a rising share. The AI-for-science cohort alone absorbed several billion in 2025: Periodic Labs raised around $300 million in a seed led by a16z and Nvidia, Lila Sciences pulled in more than $350 million across rounds, and Edison Scientific spun out of FutureHouse with backing aimed squarely at autonomous discovery agents.
Infrastructure is where the Astra datapoint cuts against intuition. A $2,000 inference bill sounds deflationary for compute demand. It is the opposite. Cheap per-result cost is what makes it rational to run thousands of speculative research agents in parallel, which is a straightforward Jevons dynamic. OpenAI's disclosed compute commitments, in the hundreds of billions across Microsoft, Oracle, CoreWeave, Broadcom and Nvidia arrangements, only pencil out if inference demand grows faster than unit cost falls. Long-horizon agents that think for days are the demand shape that justifies it.
Competitive Dynamics
Three dynamics are in play. First, credibility as a moat. After the October 2025 Erdos misfire, formal verification became a reputational requirement rather than a nicety, and labs that can produce kernel-checked artifacts get to make claims that others cannot. Second, talent gravity. Solving a Gromov problem is recruiting collateral aimed at exactly the population of mathematicians and theoretical computer scientists every lab is bidding for. Third, the closed-model problem. Because Astra is unreleased, its results function as a capability signal to investors and rivals, not as a product advantage, and open-weight prover lines from DeepSeek and Qwen continue to compress the commodity layer underneath.
Google DeepMind is the closest competitor by capability and the most disciplined about publication, having taken AlphaProof through peer review in Nature. Anthropic is running a different play, prioritizing agentic coding and enterprise reliability over research spectacle, which looks weaker in headlines and stronger in revenue quality.
Market Projections
The addressable market here is not 'mathematics'. Global academic mathematics is a rounding error in software terms. The market is verified reasoning as a component: formal methods for chip design and cryptographic protocols, automated proof obligations in safety-critical software, algorithm discovery in coding theory and lattice cryptography, and packing or scheduling optimization with provable bounds. Formal verification tooling was a low-single-digit-billion market in 2025 across EDA and software assurance; agentic research platforms plausibly compound at 40 percent or better through 2030 if certificate-backed output becomes procurement standard. The lattice cryptography result deserves specific attention from anyone tracking post-quantum migration, since NIST-standardized schemes rest on hardness assumptions that automated attack search now probes far more cheaply than before.
π Timeline & Milestones
2026 Expectations
Expect the next four months to be dominated by independent audit rather than new claims. Realistic sequence: specialists in operator algebras and geometric group theory read the 249-page manuscript against the Lean statements, and the community reaches consensus on whether the formal soficity definition matches Gromov's, probably by late autumn. Watch for arXiv postings of the individual results with named human co-authors, and for erdosproblems.com to reclassify the three entries. On the product side, a public Astra release or a research-preview tier is plausible before year end, likely gated and priced well above current reasoning models, with a per-task rather than per-token billing shape. Also expect at least one competing lab to publish a certificate-backed open-problem result before December, because the reputational cost of silence is now high. A journal accepting a paper whose primary author credit includes a model is the institutional milestone to watch.
2027-2030 Outlook
Formalization becomes the default rather than the exception for machine-assisted results, and mathlib growth shifts from human-contributed to agent-contributed, which forces a governance fight over review standards. Cost per research-grade theorem likely falls another order of magnitude, moving the constraint decisively to problem selection. By 2028 the pattern should generalize beyond proofs into any domain with a cheap verifier: chip layout, protocol security, compiler correctness, algorithm discovery. Expect the first genuine dispute where a kernel-checked result is correct yet mathematically uninteresting or subtly mis-specified, and expect that dispute to produce the field's first real standards body for formal statement review. Domains without cheap oracles, wet-lab biology in particular, will lag by years because the bottleneck is physical throughput, which is exactly why Periodic Labs and Lila are building instruments rather than only models.
Beyond 2030
The plausible end state is a two-speed research economy. Problems reducible to search plus verification get industrialized and effectively cease to be careers. Problems requiring new definitions, new frameworks, and taste about what deserves attention become the entire human contribution, and mathematical training reorients toward specification and judgment. A harder question sits underneath: if models generate proofs no human reads in full, mathematical knowledge becomes partly machine-resident, trusted through certificates rather than understood through exposition. That is a genuine epistemic change, and the discipline has not decided whether it is acceptable. Critical path dependencies: continued Lean and mathlib maturity, availability of long-horizon inference at falling cost, and enough human expert capacity to keep adjudicating statements. Remove any one and the flywheel stalls.
π° Investment Perspective
Opportunities
The cleanest exposure is inference infrastructure, because day-scale agent runs consume compute in a pattern that looks nothing like chatbot serving: long context, heavy KV cache pressure, high memory bandwidth, tolerant of latency. That profile favors Nvidia and Broadcom on silicon, Micron and SK Hynix on high-bandwidth memory, and neoclouds like CoreWeave and Nebius on capacity. Second-order: EDA and formal methods incumbents, Synopsys and Cadence, whose verification franchises get a demand tailwind if certificate-backed design becomes standard. Third: private-market access to the verification layer, where Harmonic, Axiom Math and Math Inc. sit, is where the asymmetric returns are and where public investors mostly cannot play.
Risk Factors
The single largest risk is that one or more of the ten results does not survive audit. A partial retraction would not undo the capability trend, but it would compress sentiment on research-agent narratives hard, and the October 2025 Erdos episode shows how quickly that happens. Second, Astra is unreleased, so there is zero revenue attached to this news and any multiple expansion is pure narrative. Third, the Jevons argument for compute demand is a thesis, not a fact; if agentic research proves cheap and low-volume, infrastructure capex assumptions get ugly. Fourth, concentration: OpenAI-adjacent exposure runs through a small number of private counterparties with circular financing structures. Honest limitation on my own read: I cannot verify the $2,000 figure, and no outside party can until the model or a comparable one is accessible, so treat the cost curve claim as directional.
Recommendations
Core: SMH or SOXX for semis exposure with less single-name risk than NVDA alone. QQQ and AIQ for broad AI beta. For the agent-specific theme, ARKQ and WTAI hold more of the automation and software layer. Individual names to track rather than chase: NVDA, AVGO, MU, SNPS, CDNS, CRWV, MSFT as the OpenAI proxy, and GOOGL as the only public pure-play with a peer-level formal proving stack. Strategy: avoid trading the announcement, wait for independent verification of at least the non-sofic result, and size positions to the compute-demand thesis rather than the theorem count.
π Recommended Resources
- Lean/formal verification courses
- Abstract algebra and group theory textbooks
- AI research agent tooling
Affiliate links help support AI Future Lab research.
π‘ Key Takeaways
Astra reportedly closed ten problems open for a decade or more for roughly $2,000 in inference, implying about $200 per research-grade theorem at marginal cost.
The non-sofic group construction answers a question Gromov posed in 1999 and sits in the same cohomological-rigidity lineage as the 2020 non-approximable-group work, which makes it credible rather than out of nowhere.
Lean 4 certificates published to GitHub are the structural upgrade over the October 2025 Erdos misfire; the audit burden moves from checking proofs to checking that formal statements match the real questions.
Constructions and counterexamples are where automated search plus verification has genuine advantage, because success produces a self-verifying witness. Do not extrapolate to conceptual program-building.
Cheap per-result cost increases total compute demand rather than reducing it, since it makes running thousands of speculative research agents in parallel economically rational.
The verification layer, Harmonic, Axiom Math, Math Inc., Lean FRO, is the scarce good in a world of cheap machine-generated claims, and it is mostly private-market exposure today.
Watch three things before repricing anything: independent expert confirmation of the soficity statement, arXiv postings with named co-authors, and whether OpenAI ships Astra access with per-task pricing.
π‘ Lab Test Report
π Sources & References
π€ AI Research System
Research & Analysis: Claude Opus 4.7
Infographics: Flux.1-schnell (λ‘컬)
Published: August 02, 2026
Word Count: ~2,500-3,000 words
Next Deep Dive: Next Sunday
![[Deep Dive] Bilayer Grapheneβs Magic Revealed](/content/images/size/w600/2026/08/deep_dive_thumb-11.png)
![[Deep Dive] Superconducting Hydride Mg2RhH6 Experimentally Achieved at Lower Pressure](/content/images/size/w600/2026/08/deep_dive_thumb-10.png)