AI Just Solved 10 Decades-Old Math Problems for $2,000
💡 On August 1, 2026, OpenAI announced that its next-generation model Astra solved or substantially advanced ten long-standing open problems in mathematics, with every argument formalized as machine-checkable Lean 4 code on GitHub. The total compute cost was roughly $2,000. AI is now producing verifiable mathematical results, not just plausible explanations.
- AI solves open math problems across group theory, coding theory, quantum complexity, cryptography, and combinatorics - ten results in one batch.
- Every result includes a machine-verified Lean 4 proof with zero unproven placeholders, published under an open Apache 2.0 license on GitHub.
- The most significant result: the first explicit construction of a non-sofic group, a question open since 1999.
- Total compute cost was roughly $2,000 - signalling that this level of formal reasoning is becoming accessible beyond elite institutions.
- Caveat: no formal peer review completed yet, and OpenAI selected which problems to publish with no disclosure of failure rates.
What OpenAI's Astra just announced
On August 1, 2026, OpenAI published a 249-page manuscript and a set of machine-checkable Lean 4 proof certificates on GitHub, claiming that an internal version of Astra had resolved or made substantial progress on ten open problems in mathematics and theoretical computer science. The problems span eight distinct fields: group theory, high-dimensional geometry, coding theory, operator algebras, arithmetic circuit complexity, quantum complexity, lattice cryptography, and extremal combinatorics.
The Lean files contain zero "sorry" placeholders. In formal proof assistants, a "sorry" marks an unproven step. Zero means every logical step is fully spelled out and independently verifiable by anyone who downloads the files - no trust in OpenAI required. Fields Medalist Tim Gowers said he would recommend the proofs for top-tier mathematics journals. Thomas Bloom at the University of Manchester, curator of the Erdős problems database, called the results "even more significant than the earlier unit distance counterexample."
Why does a $2,000 price tag change the equation?
OpenAI estimated the total compute cost at roughly $2,000 at GPT-5.6 Sol API rates. That figure excludes human expert time for problem selection and manuscript preparation, and covers only successful runs. Even with those caveats, it is a striking number. Historically, advancing a single open problem required years of expert time. Ten formal results spanning problems open between 27 and 48 years - for $2,000 in token costs - represents a different order of magnitude.
Well-funded university research groups can afford $2,000 per project cycle. Independent researchers and advanced graduate students could realistically run similar experiments. The economics of AI-assisted formal mathematics are shifting rapidly, not because every problem is now easy, but because the resource barrier has changed.
How do Lean 4 proofs work - and why does machine verification matter?
Lean is a programming language and proof assistant. When you write a proof in Lean, every logical step must be spelled out in machine-readable detail. If a single step is incorrect, the software rejects the entire argument. There is no room for intuitive leaps that human reviewers sometimes accept without full scrutiny. OpenAI used Lean 4.32.0 with the mathlib library for all ten results, and all files are available under an Apache 2.0 open license.
This verification model matters because the check is genuinely independent of OpenAI. Anyone can read DataCamp's analysis of the ten results and then verify the Lean proofs locally without trusting the company's claims. That is a meaningful step beyond the usual "trust the preprint" model for early research announcements.
What does this mean for you?
If you work in cybersecurity or cryptography: the lattice cryptography result is the most directly practical. Astra produced a machine-verified proof that the Closest Vector Problem remains computationally hard under standard assumptions. This underpins post-quantum encryption standards being adopted to resist quantum computers. A formally verified hardness proof, available for independent inspection, strengthens the theoretical basis for these schemes.
If you build or use AI tools professionally: this result shows AI moving from producing plausible text to producing verifiable outputs in rigorous domains. A Lean certificate is a type of trustworthy AI output that a natural-language explanation cannot replicate. This pattern - AI output combined with formal verification - is likely to reach software verification, chip design, and safety-critical systems within the next few years.
If you are a researcher, student, or educator: the Lean files are freely available and serve as educational resources for proof formalization. More broadly, the bottleneck in formal mathematics is shifting. The scarce resource is no longer computational power - it is knowing which problem is worth formalizing. Human judgment about problem selection becomes more valuable, not less.
Is this result trustworthy?
Three real limits apply. First, no formal peer review has been completed. Lean verification confirms logical consistency, but it cannot confirm that the problem was formalized correctly in the first place. An incorrect formalization that passes the checker is still wrong mathematics. The broader mathematical community is still reviewing the manuscript and the problem encodings.
Second, problem selection was controlled entirely by OpenAI. There is no disclosure of how many problems were attempted and failed, or of the total compute cost including failed runs. The $2,000 figure reflects successful inference only. This is selection bias, and naming it plainly is part of reading these results honestly.
Third, there is recent precedent for caution. In October 2025, a previous OpenAI claim about Erdős problems collapsed when independent verification proved impossible. AI critic Gary Marcus described this August 2026 batch as "remarkable" but also "vastly oversold" as evidence of general reasoning capability. The Lean proofs can be checked; broader AI capability claims cannot.
- Lean checks logical consistency - not whether the formalization captured the right mathematical question.
- OpenAI selected which problems to publish with no transparency about failure rates or total compute spent.
- Astra is not publicly available; only the published Lean certificate files can be independently accessed.
FAQ
What is a Lean proof, and why does it matter here?
Lean is a programming language that forces every step of a mathematical argument to be machine-readable. If any step is wrong, the software rejects the proof automatically. OpenAI's Lean 4 files for all ten results have zero unverified placeholders, meaning any researcher can download them and run the checker without trusting OpenAI's claims. This makes the August 2026 results more independently verifiable than a typical research preprint claim.
Can I access OpenAI's Astra right now?
No. Astra has not been publicly released. OpenAI has not announced a model card, pricing, or a ChatGPT integration date. Astra is also the first model designated for pre-release national security evaluation under the TRAINS Executive Order. Sam Altman briefed US lawmakers in late July 2026, but no public availability timeline has been confirmed as of August 2026.
Does this mean AI can now solve any math problem?
No. Astra solved problems in fields where formal proof libraries already exist - primarily discrete and algebraic mathematics with strong coverage in Lean's mathlib library. Problems requiring novel conceptual frameworks, experimental observation, or fields not yet formalized in Lean are well outside current reach. These results show impressive depth in a carefully selected domain, not general mathematical problem-solving capability.
Why does the $2,000 compute cost matter?
It matters because it resets expectations about the economics of formal mathematics research. Historically, advancing a decades-old open problem required years of expert human time. Even accounting for failed runs not included in the estimate and human setup time, $2,000 in token costs is a different order of magnitude. It suggests AI-assisted formal reasoning may soon be accessible to university research groups and well-resourced individuals, not only elite institutions.
How does this affect encryption and online security?
The lattice cryptography result is directly relevant. Astra formally proved that the Closest Vector Problem remains hard to approximate, which underpins post-quantum encryption standards like CRYSTALS-Kyber currently being adopted. A machine-verified proof of this hardness strengthens the theoretical foundation. It does not change what encryption software you should use today, but it reinforces that the mathematical basis of post-quantum encryption is formally sound.
Source(s): Forbes - OpenAI's Astra Solved Decades-Old Math Problems For $2,000 (Aug 2026); DataCamp - OpenAI's Astra Solved Ten Open Math Problems (Aug 2026)
About the author
Dao Huy (Lucas) is a professional translator with over 7 years of experience across English, Vietnamese, Chinese, and French. He follows AI and formal verification research with particular interest in how machine-generated outputs are becoming trustworthy in rigorous domains - a question as relevant to precision in technical translation as to mathematics. He has also written about Google's Gemini 3.5 Live Translate launch and the science of how CRISPR gene editing works.
He offers professional English-Vietnamese translation, technical and IP document translation, and software localization services. If you have a technical document, patent, or software product that needs precise multilingual adaptation, request a quote at daohuy.com.
Written by Dao Huy (Lucas), Vietnamese translator & localization specialist (EN · ZH · FR → Vietnamese). See translation services →