OpenAI’s Astra Solved 10 Open Math Problems: The Honest Read
💡 On August 2, 2026, OpenAI published verifiable proofs that its internal Astra model solved ten long-open math problems, from a 1999 group-theory question to Erdős catalogue entries. Each proof was formalized in Lean 4, a machine-checkable language that either accepts or rejects with no ambiguity. The line between AI research tool and AI research collaborator just shifted.
- OpenAI's Astra produced ten formally verified solutions to long-open problems in group theory, combinatorics, geometry, and lattice cryptography, costing roughly $2,000 in compute.
- Every proof was published in Lean 4, where "it compiles" means the logical chain is complete with zero steps left unproven (zero "sorry" count).
- Independent experts called the non-sofic group result "big news," and a Fields Medalist said he would recommend one result to a top journal. Peer review has not happened yet.
- Two results drew research misconduct allegations: mathematicians identified prior work not properly attributed in the sphere-packing and non-sofic group proofs.
- Astra remains private. OpenAI overstated a similar claim in October 2025. Wait for independent peer review before drawing firm conclusions.
What Did Astra Actually Solve?
Astra's demonstration of AI solving open math problems spans several branches of mathematics and theoretical computer science. The most widely noted result is the first explicit construction of a non-sofic group, answering a question Mikhail Gromov left open in 1999 when he introduced the concept of soficity in group theory. A non-sofic group cannot be approximated by finite symmetric groups in a precise technical sense, and for 27 years nobody could prove such a group existed.
Other results include a disproof of Connes's rigidity conjecture on von Neumann algebras, a proof of Ehrhart's volume conjecture in lattice geometry, three entries from Paul Erdős's open problem catalogue including problem 183 on multicolor Ramsey numbers, new sphere-packing bounds, and results in arithmetic circuit complexity, quantum parallel repetition, and lattice cryptography.
OpenAI published a 249-page manuscript of model-written reasoning walkthroughs alongside Lean 4 proof certificates on GitHub under an Apache 2.0 license. The repository reports a "sorry count" of zero: no step in any proof was left as a placeholder. One independent auditor reconstructed 8,820 of 9,007 Lean build jobs, confirming the proofs are technically reproducible.
How Does Lean 4 Verification Work?
Lean 4 is a formal proof assistant. You encode a mathematical claim in its type system and write a derivation that the Lean kernel checks step by step. The kernel enforces the rules of the underlying logic. If the proof file compiles, the derivation is internally consistent under the specified axioms. If it does not compile, it is rejected. The check is completely binary.
There is no "roughly correct" in Lean: a proof either passes or fails. A zero sorry count means no step was skipped or deferred, and the entire logical chain is present and machine-verified. This is what makes the Astra announcement different from earlier AI math claims: the proofs can be independently run and confirmed to compile.
But Lean verification does not check four things that mathematicians actually care about: whether the formal theorem statement accurately captures the original informal problem, whether the encoded objects match standard definitions in the field, whether the informal-to-formal bridges are complete, and whether the result is genuinely novel rather than a restatement of prior work. These four gaps are where most of the legitimate debate about Astra's announcement lives.
What Does This Mean for Anyone Who Does Knowledge Work?
There are three tiers of mathematical capability: mechanical calculation (computers have done this since the 1940s), solving textbook problems (AI models have handled this for several years), and original research-level reasoning - generating new results that experts had not reached. If Astra's announcement holds up under peer review, it pushes AI into the third tier for the first time in a checkable, verifiable way.
The practical implication is not that AI replaces mathematicians. Thomas Bloom, who maintains the Erdős problems catalogue, explicitly rejected that framing. He noted that the system relies on over a century of accumulated mathematical theory and was trained on the entire existing literature. What it may do is accelerate the search for proofs in areas where the conceptual framework is established but the final combination of ideas has not been found.
For anyone working in fields where formal reasoning matters - law, engineering, software, science, finance - the tools that will assist your research in five years will look different than what exists today. The research-assistant tier of AI is arriving earlier than most forecasts assumed. This joins a broader pattern of milestones redefining the boundary, including IBM's verified quantum advantage from earlier this year.
For technical translators and communication specialists, there is a specific angle: as AI takes on more formal reasoning tasks, the documents experts are asked to localize and explain will become denser and more technically precise. The human skill of translating what a formal result means into clear language for a non-specialist audience is not going away.
The Honest Limits: What the Proofs Do Not Prove
The attribution concerns are serious and deserve careful reading. Steven Miller (Yeshiva University) stated that Astra's sphere-packing proof reused arguments from his 2016 paper without proper citation, calling it "running roughshod over the work of others." Francesco Fournier-Facio (University of Cambridge) identified that the non-sofic group result combined ideas from 2016 and 2019 papers without adequate acknowledgment.
OpenAI responded by committing to corrections. But the company also cited the International Mathematical Union's Leiden Declaration - a document warning that AI companies are "using published research without consent, bypassing peer review, and threatening the integrity of proof and attribution" - and then released its findings via a company blog post rather than a peer-reviewed journal. That is the exact practice the Declaration warns against.
Independent auditors found that only one result could be confirmed as genuinely novel at the time of reporting. And OpenAI has a relevant track record: in October 2025, a company VP claimed GPT-5 had solved ten Erdős problems. Bloom called that claim "a dramatic misrepresentation" - the model had only rediscovered already-published results, and the post was later deleted. Read the full misconduct allegations before deciding how much weight to give this announcement.
What Should You Watch For in the Coming Months?
The most important signal will be formal peer review. If one or more of the ten results is accepted by a top mathematics journal, the announcement crosses from "promising claim" to "documented advance." Fields Medalist Timothy Gowers reportedly said he would recommend one result to a top journal without hesitation. If that nomination proceeds and succeeds, it is the strongest possible confirmation.
Astra will eventually be released publicly. When it is, independent researchers can attempt to reproduce and extend the results - the normal scientific process, delayed by the private-model release strategy. The $2,000 compute cost for ten proofs suggests that when the model is available, exploring additional open problems at scale becomes practically feasible for well-funded research groups.
For the broader question of AI solving open math problems, this announcement adds to a growing record that includes AlphaProof's silver-medal performance at the 2024 International Mathematical Olympiad. The trend line is real, even if individual announcements deserve scrutiny. The pace of progress is faster than most people expected three years ago.
FAQ
Does this prove AI can now do original mathematical research?
Cautiously, yes - for problems where the conceptual tools are established and the challenge is combining them correctly. Astra found valid combinations in formally verifiable ways. But peer review of these specific results is still pending, and the attribution disputes mean not all ten claims stand as fully original contributions yet.
What exactly does a Lean 4 proof confirm?
A Lean proof confirms internal logical consistency: the derivation follows the rules of the logic system without gaps. It does not confirm that the formal theorem matches the original informal problem, that definitions are standard in the field, or that the result is genuinely new. Lean verification is a necessary but not sufficient condition for mathematical acceptance.
Are the attribution concerns serious?
Yes. Two researchers identified specific results where prior work was not properly credited. OpenAI committed to corrections. Mathematical authorship carries accountability for definitions, citations, and error correction that a model cannot independently carry. The International Mathematical Union's Leiden Declaration frames attribution and peer review as systemic concerns for AI and mathematics broadly.
Can I use Astra to solve problems of my own?
Not yet. Astra is still a private, internal model with no public release date or pricing announced. When it becomes available, it will likely be accessible through the OpenAI API, but the deployment model has not been decided. OpenAI may fold it into GPT-6 or release it separately.
Did Astra solve any Millennium Prize Problems?
No. OpenAI researcher Noam Brown confirmed the team tried but did not succeed. Those seven problems, including P vs NP and the Riemann Hypothesis, remain open. They are in a different difficulty category from the ten problems Astra tackled, and the compute cost to attempt them would be far higher than $2,000.
Source(s): SiliconAngle (Aug 2026); AI Weekly (Aug 2026); The Decoder (Aug 2026)
About the author
Dao Huy (Lucas) is a professional translator working across English, Vietnamese, Chinese, and French, with over seven years of experience in technical and IP translation. He follows the frontier of AI and science out of genuine curiosity, particularly where formal language, verification, and clear communication intersect. Proof languages like Lean are a fascinating case: precise enough for machines to check, yet encoding ideas that human researchers spent decades pursuing in informal notation.
He offers English-Vietnamese translation, technical document localization, and patent and IP translation services. If you have a research document, technical specification, or IP filing that needs precise, context-aware translation, get a quote at daohuy.com.
Written by Dao Huy (Lucas), Vietnamese translator & localization specialist (EN · ZH · FR → Vietnamese). See translation services →