OpenAI’s Unreleased Astra Model Cracks Ten Long-Standing Mathematics Problems

Artificial intelligence has moved beyond solving examination questions and benchmark puzzles.
OpenAI says an unreleased model from its next major model family has produced ten significant advances across mathematics and theoretical computer science—some involving problems that had resisted researchers for decades.
The internal model, called Astra, generated results spanning geometry, coding theory, group theory, operator algebras, quantum complexity, cryptography and combinatorics.
OpenAI published the work on 1 August 2026 under the title “Ten Advances in Mathematics and Theoretical Computer Science.”
The company says the model resolved some open problems outright and made substantial progress on others. Each argument was subsequently prepared into a manuscript and formalised in Lean, a proof-assistant system capable of checking whether every logical step is valid. �
OpenAI
What Astra Achieved
OpenAI listed ten areas in which Astra produced new mathematical results.
1. High-Dimensional Sphere Packing
Astra developed new upper bounds on the density of sphere packing in high-dimensional spaces.
Sphere packing asks how efficiently identical spheres can be arranged without overlapping.
The problem has connections to coding theory, communications, geometry and information storage.
2. Binary and Spherical Codes
The model produced exponentially improved bounds on the maximum size of binary codes at prescribed minimum distances, with related results for spherical codes.
These problems help determine how information can be transmitted reliably even when noise or errors are present.
3. Non-Sofic Groups
Astra constructed an example proving that non-sofic groups exist.
For decades, mathematicians had asked whether every group could be approximated through finite symmetric structures.
Astra’s construction showed that the answer is no. �
OpenAI
4. Connes’s Rigidity Conjecture
The system produced a disproof of Alain Connes’s rigidity conjecture.
The conjecture concerned whether certain mathematical groups are uniquely determined by the von Neumann algebras associated with them.
Astra found that this proposed uniqueness does not always hold. �
OpenAI
5. Arithmetic Circuit Complexity
The model developed new lower bounds for computing the permanent using arithmetic circuits and formulas.
This area investigates the minimum computational resources required to solve difficult algebraic problems.
Such results contribute to broader research on what computers can calculate efficiently.
6. Quantum Parallel Repetition
Astra established an exponential parallel-repetition theorem for general two-player quantum games.
Parallel-repetition results explore what happens when the same game or verification challenge is repeated many times.
They are important in complexity theory, cryptography and the study of quantum information.
7. The Closest-Vector Problem
The model produced polynomial-factor hardness results for approximating the closest-vector problem.
This is a foundational problem involving mathematical lattices and has major relevance to post-quantum cryptography.
8. Ehrhart’s Volume Conjecture
Astra determined the maximum possible volume of certain convex bodies in every dimension under specific lattice conditions.
This resolved a long-standing geometric conjecture associated with Eugène Ehrhart. �
OpenAI
9. Multicolour Ramsey Numbers
The system established a superexponential lower bound for multicolour triangle Ramsey numbers.
OpenAI says this resolves Erdős problem 183.
Ramsey theory studies how order inevitably appears inside sufficiently large structures.
10. Extremal Graph Theory
Astra produced results resolving the compactness and degeneracy conjectures in extremal graph theory.
OpenAI says these results settle Erdős problems 146 and 180. �
OpenAI
How Were the Proofs Checked?
This is one of the most important parts of the announcement.
Large language models can produce convincing but incorrect reasoning. A mathematical result therefore cannot be accepted merely because the explanation sounds sophisticated.
OpenAI says Astra formalised each argument into a Lean certificate.
Lean is a formal proof system. It does not evaluate whether a proof feels persuasive. Every step must follow strictly from established definitions and logical rules.
If the formalisation succeeds, it gives much stronger assurance that the proof contains no hidden logical gap.
However, formal verification does not answer every scientific question.
Mathematicians must still determine:
Whether the result is genuinely new
Whether the definitions match the intended problem
How important the result is
Whether the proof introduces useful ideas
How the work connects with existing literature
Whether independent experts can reproduce and understand it
A machine-checked proof can establish correctness.
The mathematical community must establish significance.
How Much Did the Work Cost?
OpenAI says the total number of tokens required to find all ten results would cost approximately $2,000 using Sol API prices. �
OpenAI
That figure may prove almost as disruptive as the results themselves.
Advanced mathematical research normally requires years of specialised education, institutional funding, collaboration and sustained human attention.
If a general-purpose AI system can explore major open problems for thousands rather than millions of dollars, access to research-level discovery could broaden dramatically.
A university department, small research team or independent mathematician might eventually use models to explore questions that previously demanded far larger resources.
The economic barrier to experimentation could fall.
The intellectual barrier will remain—but it may change form.
Did Astra Work Entirely Alone?
OpenAI says the mathematical arguments themselves were generated by the model.
Human researchers then helped prepare the outputs into manuscripts, reviewed the work, formalised the proofs and accepted responsibility for correctness.
OpenAI explicitly said it would be misleading to claim purely human authorship for arguments generated entirely by an AI system. �
OpenAI
This creates a major change in academic authorship.
Traditionally, researchers receive credit because they developed the ideas, wrote the proof and accepted responsibility for the result.
When an AI generates the central argument, humans may still:
Select the problem
Design the prompts
Interpret the result
Check originality
Repair presentation
Formalise the proof
Explain its importance
Take legal and professional responsibility
But the creative mathematical contribution may no longer be entirely human.
Science will need a new vocabulary for that collaboration.
Anthropic Enters the Story
Within approximately 24 hours, Anthropic researcher Levent Alpöge claimed that Claude Fable reproduced five of Astra’s ten results using generic prompting and no internet access.
The reported results included work in arithmetic-circuit complexity, quantum parallel repetition and the closest-vector problem. �
The Rundown AI +1
This matters for two reasons.
First, it suggests the breakthroughs may not depend on a single secret trick available only inside OpenAI.
Second, it raises the possibility that multiple frontier systems are crossing the threshold into genuine research-level mathematics at roughly the same time.
However, the claim should remain carefully qualified.
At the time of reporting, Anthropic had not published a complete formal paper independently verifying all five reproductions. �
Moneycontrol
Could an AI Receive a Fields Medal?
The Fields Medal is awarded to human mathematicians under the age of 40 for outstanding contributions.
An AI model does not meet those traditional criteria.
It has no age in the human sense, no legal identity, no academic career and no personal responsibility.
Therefore, Astra itself would not be eligible under the conventional structure of the prize.
But the philosophical question is still important.
If an AI produces mathematics judged worthy of the highest honour, who receives the recognition?
Possible answers include:
The human research team
The company that created the model
The mathematicians who selected the problem
The engineers who trained the system
Nobody under existing prize rules
A new category created specifically for machine-generated discovery
The debate is not really about medals.
It is about authorship, intellectual labour and whether society is ready to recognise non-human systems as contributors to knowledge.
Why the Story Went Viral
The story contains several powerful ingredients.
It involves an unreleased model.
It involves decade-spanning mysteries.
The proofs were reportedly machine-verified.
The total inference cost was surprisingly low.
And a rival AI laboratory quickly claimed it could reproduce half of the achievements.
This is no longer a simple benchmark competition.
The models are not merely being asked to answer questions with known solutions.
They are being asked to create knowledge that did not previously exist.
That is a different category of capability.
The Bigger Scientific Implication
Mathematics is an unusually strong testing ground for AI discovery because proofs can be checked precisely.
A drug-discovery model may suggest a promising molecule, but researchers still need laboratory testing, clinical trials and years of evidence.
A mathematical proof can sometimes be verified much more directly.
If the logic is correct and the result is genuinely new, the contribution is real.
This may make mathematics the first major scientific discipline where AI-generated discovery becomes routine.
The same research pattern could later expand into:
Materials science
Theoretical physics
Chemistry
Drug discovery
Engineering design
Climate modelling
Energy systems
Computer-security research
In those fields, however, verification will often require physical experiments rather than formal logic.
AI may generate scientific possibilities faster than laboratories can test them.
Mary Chuks’ Perspective
For years, people asked whether AI could think.
This announcement sharpens the question.
Can AI contribute something humanity did not know before?
If the Astra results survive full independent mathematical review, the answer is increasingly yes.
But intelligence alone does not complete science.
Human beings still choose which questions matter.
Humans interpret why an answer is important.
Humans decide how knowledge should be used.
And humans remain responsible for ensuring that speed does not outrun understanding.
The best future is not mathematics without mathematicians.
It is mathematics in which human curiosity directs machines capable of exploring intellectual territory at unprecedented scale.
AI becomes the explorer.
Humanity must remain the navigator.
Practical Takeaways
For researchers and institutions:
Treat AI-generated proofs as research candidates, not unquestionable truth.
Use formal verification wherever possible.
Check the literature carefully for prior results.
Preserve full records of model prompts and outputs.
Disclose the role of AI honestly.
Invite independent experts to review important claims.
Develop new authorship and attribution standards.
Provide broad research access rather than concentrating capability inside a few companies.
Distinguish full resolutions from partial advances.
Prepare universities for AI-assisted original research.
Conclusion
OpenAI’s Astra announcement may represent a turning point.
The model did not simply solve difficult textbook questions.
According to OpenAI, it generated new mathematical arguments addressing ten long-standing research problems, formalised them for machine checking and did so at a surprisingly low inference cost.
Independent review will determine how transformative each result ultimately proves to be.
But the direction is already difficult to ignore.
Artificial intelligence is moving from learning humanity’s existing knowledge to helping create the next part of it.
And that changes the meaning of the word researcher.
Original Sources and Further Reading
Primary source: OpenAI, “Ten Advances in Mathematics and Theoretical Computer Science,” published 1 August 2026. The publication lists all ten results, explains that an internal Astra model generated the arguments and links to manuscripts, reasoning walkthroughs and Lean certificates. �
OpenAI
Original newsletter summary: The Rundown AI, “OpenAI’s ‘Astra’ Solves 10 Long-Standing Math Problems,” published 3 August 2026. �
The Rundown AI
Additional independent reporting: Moneycontrol, “OpenAI Claims Its Next-Generation AI Model Astra Solved 10 Complex Math Problems; Anthropic Says Claude Fable Solved Five,” published 3 August 2026. �
Moneycontrol