OpenAI announced on 1 August 2026 that an internal version of Astra, its next model family, has produced solutions to ten open problems in mathematics, spanning group theory, von Neumann algebras, sphere packing, coding theory, quantum complexity, lattice cryptography and extremal combinatorics, at a stated cost of roughly $2,000 in API tokens.1 The Decoder 2026-08-01 Internal Astra version resolved ten previously unsolved problems for roughly $2,000 in API tokens at Sol rates; problems stalled for at least a decade; Thomas Bloom called the results big news and more significant than the May unit distance counterexample; Noam Brown acknowledged failed Millennium Prize attempts; OpenAI targets a research intern level system by September 2026 and an autonomous AI researcher by March 2028; compounding errors in long workflows acknowledged as a weakness. Open source The results shipped as a 249 page manuscript with machine checkable Lean 4 certificates published on GitHub, and they follow a May result in which an OpenAI model disproved the Erdős unit distance conjecture, an 80 year old question in discrete geometry.2 The Next Web 2026-08-01 The ten results: first non-sofic group construction (open since Gromov, 1999), disproof of Connes's rigidity conjecture, proof of Ehrhart's volume conjecture, three Erdős problems including number 183, first improvement to the sphere packing upper bound since 1978, parallel repetition for two player quantum games, circuit lower bounds for the permanent; Lean 4 certificates on GitHub and a 249 page manuscript; Gowers quote on Annals of Mathematics; the June 2026 Leiden Declaration endorsed by the IMU criticized results by press release. Open source 3 Markman Capital Insight 2026-08-03 $2,000 covers successful runs only; problem selection may have been curated; OpenAI staff assisted formalization; independent reproduction impossible, only verification; Gary Marcus called the results likely vastly oversold; formal verification industries (chip design, cryptography, safety critical software) framed as natural adopters. Open source The stake is whether frontier labs can now buy publishable mathematics with compute, and how anyone outside the lab is supposed to know. We assess with high confidence that the durable shift here is the verification format, not the headline count: a formal certificate converts a press release claim into something a stranger can check, and that is new for announcements of this kind. We assess with moderate confidence that the $2,000 figure understates the true cost by a wide margin, because by the reporting it counts only the successful runs.3 Markman Capital Insight 2026-08-03 $2,000 covers successful runs only; problem selection may have been curated; OpenAI staff assisted formalization; independent reproduction impossible, only verification; Gary Marcus called the results likely vastly oversold; formal verification industries (chip design, cryptography, safety critical software) framed as natural adopters. Open source

What was actually solved

The list is specific enough to audit. It includes the first explicit construction of a non-sofic group, closing a question open since Mikhail Gromov introduced soficity in 1999; a disproof of Connes's rigidity conjecture for von Neumann algebras; a proof of Ehrhart's volume conjecture; three problems from the Erdős catalogue, including number 183 on multicolored Ramsey numbers; the first improvement to the general sphere packing upper bound since 1978; a parallel repetition theorem for two player quantum games; and new circuit lower bounds for computing the permanent.2 The Next Web 2026-08-01 The ten results: first non-sofic group construction (open since Gromov, 1999), disproof of Connes's rigidity conjecture, proof of Ehrhart's volume conjecture, three Erdős problems including number 183, first improvement to the sphere packing upper bound since 1978, parallel repetition for two player quantum games, circuit lower bounds for the permanent; Lean 4 certificates on GitHub and a 249 page manuscript; Gowers quote on Annals of Mathematics; the June 2026 Leiden Declaration endorsed by the IMU criticized results by press release. Open source These are not benchmark items. By the reporting, mathematicians had made no progress on them for at least a decade, and on most for much longer.1 The Decoder 2026-08-01 Internal Astra version resolved ten previously unsolved problems for roughly $2,000 in API tokens at Sol rates; problems stalled for at least a decade; Thomas Bloom called the results big news and more significant than the May unit distance counterexample; Noam Brown acknowledged failed Millennium Prize attempts; OpenAI targets a research intern level system by September 2026 and an autonomous AI researcher by March 2028; compounding errors in long workflows acknowledged as a weakness. Open source

The reception from working mathematicians was warmer than skeptics might expect. Thomas Bloom of the University of Manchester called the results "big news" and judged them more significant than the May unit distance counterexample, while rejecting the framing that the model replaces mathematicians, since it draws on more than a century of accumulated theory.1 The Decoder 2026-08-01 Internal Astra version resolved ten previously unsolved problems for roughly $2,000 in API tokens at Sol rates; problems stalled for at least a decade; Thomas Bloom called the results big news and more significant than the May unit distance counterexample; Noam Brown acknowledged failed Millennium Prize attempts; OpenAI targets a research intern level system by September 2026 and an autonomous AI researcher by March 2028; compounding errors in long workflows acknowledged as a weakness. Open source That is a fact worth holding next to the hype cycle: the enthusiasm on record comes from people whose careers depend on being right about what counts as a real theorem.

The May precedent built the credibility this drop is spending

The August announcement lands on ground prepared in May, when an OpenAI model produced a counterexample to the unit distance conjecture, a problem that had resisted attack since 1946. Timothy Gowers, a Fields Medalist, said he would have recommended that proof "for publication in Annals of Mathematics without hesitation."2 The Next Web 2026-08-01 The ten results: first non-sofic group construction (open since Gromov, 1999), disproof of Connes's rigidity conjecture, proof of Ehrhart's volume conjecture, three Erdős problems including number 183, first improvement to the sphere packing upper bound since 1978, parallel repetition for two player quantum games, circuit lower bounds for the permanent; Lean 4 certificates on GitHub and a 249 page manuscript; Gowers quote on Annals of Mathematics; the June 2026 Leiden Declaration endorsed by the IMU criticized results by press release. Open source More telling than the quote is what nine mathematicians, Gowers and Noga Alon among them, did next: they wrote a human verified expository digest of the machine's argument and posted it to arXiv on 20 May 2026, noting that the proof relies on ideas attributable in retrospect to Ellenberg-Venkatesh, Golod-Shafarevich and Hajir-Maire-Ramakrishna.4 arXiv 2026-05-20 Nine author expository note, W. T. Gowers and Noga Alon among the authors, presenting a short, digested, human verified version of the OpenAI generated counterexample to the Erdős unit distance conjecture, with the argument attributable in retrospect to Ellenberg-Venkatesh, Golod-Shafarevich and Hajir-Maire-Ramakrishna ideas. Open source

That paper is the template for how this kind of result gets absorbed: the machine produces the argument, humans translate it into the literature, and the literature decides what it was made of. The August drop shortcuts part of that loop with Lean 4 certificates, which let anyone verify correctness mechanically without access to the model that produced it.2 The Next Web 2026-08-01 The ten results: first non-sofic group construction (open since Gromov, 1999), disproof of Connes's rigidity conjecture, proof of Ehrhart's volume conjecture, three Erdős problems including number 183, first improvement to the sphere packing upper bound since 1978, parallel repetition for two player quantum games, circuit lower bounds for the permanent; Lean 4 certificates on GitHub and a 249 page manuscript; Gowers quote on Annals of Mathematics; the June 2026 Leiden Declaration endorsed by the IMU criticized results by press release. Open source 3 Markman Capital Insight 2026-08-03 $2,000 covers successful runs only; problem selection may have been curated; OpenAI staff assisted formalization; independent reproduction impossible, only verification; Gary Marcus called the results likely vastly oversold; formal verification industries (chip design, cryptography, safety critical software) framed as natural adopters. Open source Verification is not reproduction: no outside party can rerun Astra, so the certificates prove the theorems are true, not that the discovery process is what OpenAI says it is.3 Markman Capital Insight 2026-08-03 $2,000 covers successful runs only; problem selection may have been curated; OpenAI staff assisted formalization; independent reproduction impossible, only verification; Gary Marcus called the results likely vastly oversold; formal verification industries (chip design, cryptography, safety critical software) framed as natural adopters. Open source That distinction is where most of the open questions below live.

Who gains and who loses

OpenAI gains the most, and deliberately. Astra has no release date; the math drop is its introduction, positioned against the company's stated targets of a research intern level system by September 2026 and a fully autonomous AI researcher by March 2028.1 The Decoder 2026-08-01 Internal Astra version resolved ten previously unsolved problems for roughly $2,000 in API tokens at Sol rates; problems stalled for at least a decade; Thomas Bloom called the results big news and more significant than the May unit distance counterexample; Noam Brown acknowledged failed Millennium Prize attempts; OpenAI targets a research intern level system by September 2026 and an autonomous AI researcher by March 2028; compounding errors in long workflows acknowledged as a weakness. Open source Ten certified theorems is a stronger recruiting and fundraising artifact than any benchmark table, precisely because it cannot be dismissed as overfitting to a test set.

Industries that already run on formal verification gain a concrete signal: chip design, cryptography and safety critical software are named in the investor facing coverage as the natural first buyers of a system whose output arrives machine checkable.3 Markman Capital Insight 2026-08-03 $2,000 covers successful runs only; problem selection may have been curated; OpenAI staff assisted formalization; independent reproduction impossible, only verification; Gary Marcus called the results likely vastly oversold; formal verification industries (chip design, cryptography, safety critical software) framed as natural adopters. Open source Working mathematicians gain a tool and lose a monopoly on a specific kind of labor; Bloom's position, that this is an instrument built from their own century of theory rather than a replacement, is the optimistic version of that trade.1 The Decoder 2026-08-01 Internal Astra version resolved ten previously unsolved problems for roughly $2,000 in API tokens at Sol rates; problems stalled for at least a decade; Thomas Bloom called the results big news and more significant than the May unit distance counterexample; Noam Brown acknowledged failed Millennium Prize attempts; OpenAI targets a research intern level system by September 2026 and an autonomous AI researcher by March 2028; compounding errors in long workflows acknowledged as a weakness. Open source

The clearest loser is the traditional gatekeeping process. The Leiden Declaration of June 2026, endorsed by the International Mathematical Union, specifically criticized companies announcing results by press release rather than peer review, and this announcement did exactly that, with certificates standing in for referees.2 The Next Web 2026-08-01 The ten results: first non-sofic group construction (open since Gromov, 1999), disproof of Connes's rigidity conjecture, proof of Ehrhart's volume conjecture, three Erdős problems including number 183, first improvement to the sphere packing upper bound since 1978, parallel repetition for two player quantum games, circuit lower bounds for the permanent; Lean 4 certificates on GitHub and a 249 page manuscript; Gowers quote on Annals of Mathematics; the June 2026 Leiden Declaration endorsed by the IMU criticized results by press release. Open source Rival labs lose narrative ground: whatever their internal results, none has published a comparable certified batch. And the Millennium Prize Problems are a named casualty of honesty here: Noam Brown acknowledged OpenAI tried them and failed, which usefully bounds what this generation of systems can do.1 The Decoder 2026-08-01 Internal Astra version resolved ten previously unsolved problems for roughly $2,000 in API tokens at Sol rates; problems stalled for at least a decade; Thomas Bloom called the results big news and more significant than the May unit distance counterexample; Noam Brown acknowledged failed Millennium Prize attempts; OpenAI targets a research intern level system by September 2026 and an autonomous AI researcher by March 2028; compounding errors in long workflows acknowledged as a weakness. Open source

The counter-case

The strongest argument against the significance of the drop is selection. OpenAI chose which problems to attempt, which successes to publish and which failures to omit; the $2,000 covers successful runs only, staff assisted with the formalization, and no independent party can reproduce the discovery process.3 Markman Capital Insight 2026-08-03 $2,000 covers successful runs only; problem selection may have been curated; OpenAI staff assisted formalization; independent reproduction impossible, only verification; Gary Marcus called the results likely vastly oversold; formal verification industries (chip design, cryptography, safety critical software) framed as natural adopters. Open source Gary Marcus called the results likely "vastly oversold," and some specialists quoted in the same reporting expect only a few of the ten to be genuinely surprising once examined.3 Markman Capital Insight 2026-08-03 $2,000 covers successful runs only; problem selection may have been curated; OpenAI staff assisted formalization; independent reproduction impossible, only verification; Gary Marcus called the results likely vastly oversold; formal verification industries (chip design, cryptography, safety critical software) framed as natural adopters. Open source On the model side, OpenAI itself acknowledges that compounding errors over long agentic workflows remain a major weakness, which is the exact capability Astra is supposed to demonstrate.1 The Decoder 2026-08-01 Internal Astra version resolved ten previously unsolved problems for roughly $2,000 in API tokens at Sol rates; problems stalled for at least a decade; Thomas Bloom called the results big news and more significant than the May unit distance counterexample; Noam Brown acknowledged failed Millennium Prize attempts; OpenAI targets a research intern level system by September 2026 and an autonomous AI researcher by March 2028; compounding errors in long workflows acknowledged as a weakness. Open source

For the thesis of this brief to fail, the certificates would have to matter less than they appear: if the mathematical community concludes that most of the ten are incremental consequences of known techniques, formally checkable but intellectually thin, then the verification format is packaging on ordinary work, and the May counterexample remains the only result of consequence. That outcome is live. We assess it as less likely than not, given Bloom's on the record judgment that the batch exceeds the May result, but the review has barely begun.1 The Decoder 2026-08-01 Internal Astra version resolved ten previously unsolved problems for roughly $2,000 in API tokens at Sol rates; problems stalled for at least a decade; Thomas Bloom called the results big news and more significant than the May unit distance counterexample; Noam Brown acknowledged failed Millennium Prize attempts; OpenAI targets a research intern level system by September 2026 and an autonomous AI researcher by March 2028; compounding errors in long workflows acknowledged as a weakness. Open source

What to watch

  • Independent audits of the certificates. If outside groups confirm the Lean 4 proofs on GitHub check cleanly by the end of August 2026, the correctness question closes; any certificate that fails would be far bigger news than the announcement.2 The Next Web 2026-08-01 The ten results: first non-sofic group construction (open since Gromov, 1999), disproof of Connes's rigidity conjecture, proof of Ehrhart's volume conjecture, three Erdős problems including number 183, first improvement to the sphere packing upper bound since 1978, parallel repetition for two player quantum games, circuit lower bounds for the permanent; Lean 4 certificates on GitHub and a 249 page manuscript; Gowers quote on Annals of Mathematics; the June 2026 Leiden Declaration endorsed by the IMU criticized results by press release. Open source
  • Expository papers on the ten. Watch for a repeat of the May pattern: human authored digests of the non-sofic group construction or the sphere packing bound on arXiv by the fourth quarter of 2026 would signal the community judges them deep rather than mechanical.4 arXiv 2026-05-20 Nine author expository note, W. T. Gowers and Noga Alon among the authors, presenting a short, digested, human verified version of the OpenAI generated counterexample to the Erdős unit distance conjecture, with the argument attributable in retrospect to Ellenberg-Venkatesh, Golod-Shafarevich and Hajir-Maire-Ramakrishna ideas. Open source
  • The September 2026 intern milestone. OpenAI set itself a dated target of a research intern level system by September 2026; whether it claims that milestone, and with what evidence, calibrates how much to trust the March 2028 autonomous researcher target.1 The Decoder 2026-08-01 Internal Astra version resolved ten previously unsolved problems for roughly $2,000 in API tokens at Sol rates; problems stalled for at least a decade; Thomas Bloom called the results big news and more significant than the May unit distance counterexample; Noam Brown acknowledged failed Millennium Prize attempts; OpenAI targets a research intern level system by September 2026 and an autonomous AI researcher by March 2028; compounding errors in long workflows acknowledged as a weakness. Open source
  • Astra's shipped form and price. When Astra releases, watch whether paying customers can elicit comparable results at anything near the stated token economics, or whether open problem mathematics remains an internal, staff assisted production.3 Markman Capital Insight 2026-08-03 $2,000 covers successful runs only; problem selection may have been curated; OpenAI staff assisted formalization; independent reproduction impossible, only verification; Gary Marcus called the results likely vastly oversold; formal verification industries (chip design, cryptography, safety critical software) framed as natural adopters. Open source
  • A certified answer from a rival lab. If Google DeepMind or Anthropic publishes its own machine checkable open problem results by early 2027, certificates become the standard of evidence for capability claims, which is the lasting change this drop points to.

The theorems will be absorbed into the literature with or without OpenAI's name attached. What the next six months decide is the evidentiary standard: whether frontier capability claims now arrive with proofs a stranger can check, or whether this batch remains a one time purchase of credibility.