# OpenAI Astra Model Achieves Ten Advances in Mathematics and Theoretical Computer Science

> An internal version of OpenAI's Astra model has generated formal proofs for ten longstanding problems in mathematics and theoretical computer science, including the disproof of Connes’s rigidity conjecture, with all proofs formalized in Lean and released publicly at low cost.

*Published 2026-08-01 · By The Intel Desk*

OpenAI Astra is an internal version of the next major model from OpenAI that produced ten significant advances in mathematics and theoretical computer science.

OpenAI has revealed that an internal version of its Astra model has made substantial progress on challenging problems in mathematics and theoretical computer science that have remained unsolved for many years. The model has delivered ten notable results, each accompanied by formal proofs in the Lean theorem prover that can be independently verified. These accomplishments highlight the growing capability of advanced AI systems to contribute to pure research fields that have traditionally required human expertise over extended periods of time and deep domain knowledge. The announcement includes the release of the Lean certificates on a dedicated GitHub repository, along with detailed chain of thought walkthroughs that explain the reasoning process used by the model for each problem. The entire process of generating the solutions consumed tokens that would cost roughly two thousand dollars when using Sol API rates, according to the company report. This combination of significant results and low relative cost underscores the potential for AI to accelerate progress in theoretical disciplines.

## What background and context exist for AI contributions to mathematics?

Mathematics and theoretical computer science have long served as benchmarks for artificial intelligence systems due to their requirement for precise logical reasoning and creative insight into abstract structures. Previous efforts by various research groups have shown that AI can assist in conjecture generation and verification in limited domains, but the scale and significance of the problems tackled by Astra represent a notable step forward in the field. The release of the proofs allows the broader community to examine the reasoning steps taken by the model in a transparent manner. This transparency is important for understanding how the model arrived at its conclusions in fields like quantum complexity and lattice cryptography where the concepts are highly technical. The problems span a wide range of disciplines, indicating the versatility of the approach taken by the Astra model in handling diverse mathematical structures and theoretical challenges.

The context also includes the ongoing efforts in the AI community to integrate formal verification tools like Lean into model training and inference pipelines to ensure reliability. By producing certificates that can be checked independently, the model provides evidence that its outputs are not only plausible but rigorously correct according to the standards of the mathematical community. This is particularly valuable in areas such as arithmetic circuit complexity and extremal combinatorics where errors can be subtle and difficult to detect without formal methods. The announcement builds on earlier work in AI for math but extends it to conjectures that have resisted solution for many years despite the efforts of human researchers. Researchers in the field have noted the potential for such tools to accelerate discovery in pure mathematics by providing new avenues for exploration.

The integration of AI into mathematics is part of a broader trend where computational tools are used to explore areas that were previously inaccessible due to the complexity of the calculations involved. The Astra model has shown that it can navigate these complexities effectively, producing results that align with the expectations of the mathematical community. This is evidenced by the formal verification process that confirms the correctness of each proof without relying on human intervention for the final check. The low compute cost also suggests that the model is efficient in its use of resources, which is an important factor for scaling such capabilities to even more challenging problems in the future.

## What new results in detail has the Astra model produced?

The ten advances cover several key areas as outlined in the OpenAI report on the achievements. In high-dimensional sphere packing, the model provided better bounds that could influence packing efficiency in higher dimensions and related applications. For coding theory, new constructions or bounds were established that may have implications for the development of error-correcting codes used in communications. In arithmetic circuit complexity, advancements were made in understanding the resources needed for certain computations and their limitations. The group theory result on non-sofic groups establishes their existence, which has implications for the classification of groups and their properties. The disproof of Connes’s rigidity conjecture challenges previous assumptions in operator algebras regarding the uniqueness of groups based on their von Neumann algebras and opens new questions in the field.

Additional results include improvements in quantum complexity, potentially affecting how quantum systems are modeled computationally and their computational power is assessed. In lattice cryptography, the model contributed to better understanding of security assumptions used in post-quantum cryptographic systems that are designed to resist quantum attacks. Extremal combinatorics saw new bounds on monochromatic triangles in multicolored graphs, which relates to problems in Ramsey theory and graph coloring. These results are not isolated but demonstrate a consistent ability to tackle problems across the spectrum of theoretical computer science and mathematics. The release includes CoT walkthroughs that detail the chain of thought used by the model in arriving at each solution, providing valuable data for further research into AI reasoning.

## What are the technical specifics of the Lean formalizations and compute costs?

Each of the ten arguments was formalized by the model in a Lean certificate, ensuring that the proofs are machine-checkable and free from logical gaps that could undermine their validity. The GitHub repository contains the Lean 4 files for each result, allowing independent verification by mathematicians and computer scientists around the world. The reasoning walkthroughs provide insight into the step-by-step process the model followed, which can be useful for training future models on similar tasks and improving their performance. The total number of tokens required for generating these solutions amounted to a cost of roughly two thousand dollars when priced at Sol API rates. This efficiency in compute usage is noteworthy given the complexity of the problems addressed and the depth of the proofs produced.

Lean is a proof assistant that allows for the formalization of mathematical statements and proofs in a way that can be automatically verified by a computer system. By producing these certificates, the Astra model has demonstrated not only the ability to find solutions but also to express them in a formal language that eliminates ambiguity and ensures correctness. This dual capability of discovery and formalization is critical for advancing the field of automated theorem proving and integrating AI into the mathematical workflow. The specific file for the Connes rigidity conjecture is named ConnesRigidity.lean, as noted in the repository documentation, and similar files exist for the other results. The approach combines the generative power of the model with the verification power of the proof assistant to produce reliable outputs.

Overview of key advances from the Astra model in various mathematical fieldsFieldAdvanceDetailsOperator AlgebrasDisproof of Connes’s rigidity conjectureCounterexample showing groups not uniquely determined by von Neumann algebrasGroup TheoryExistence of non-sofic groupsConstruction establishing that non-sofic groups existHigh-dimensional sphere packingBetter boundsImproved bounds for high dimensional sphere packingCircuit complexityAdvancements in boundsBetter bounds for circuit complexityExtremal combinatoricsMonochromatic trianglesBounds for monochromatic triangles in multicolored graphs

## What are the market and stakeholder implications of these advances?

These results have implications for how stakeholders view the capabilities of frontier models in research-oriented tasks that require high levels of precision and creativity. Companies and research institutions may increase investment in AI systems that can handle formal reasoning, potentially leading to new tools for mathematicians and theoretical computer scientists. The low cost relative to the significance of the results suggests that such models could become accessible for academic use in the future if the technology continues to improve. Stakeholders in the AI industry will likely monitor how these capabilities evolve in subsequent model releases from OpenAI and its competitors in the space. The public release of the proofs encourages collaboration and further validation by the academic community, which can lead to additional discoveries based on these foundations.

For the market, this could accelerate the adoption of AI in scientific discovery, moving beyond generative tasks to verifiable contributions in hard sciences and theoretical fields. The involvement of researchers like Sébastien Bubeck and Greg Brockman in highlighting these results underscores the internal priority placed on these capabilities within the organization. As models like Astra are developed further, the ability to solve open problems could become a key differentiator in the competitive landscape of frontier AI development. This may also prompt discussions on the ethical use of AI in research and the attribution of discoveries made with AI assistance to ensure proper credit and accountability in the scientific process.

## What expert reactions have emerged regarding the Astra model results?

Reactions from within OpenAI have been positive, with key figures sharing details on social media platforms about the nature of the achievements. The results are described as beautiful and wide-ranging, covering areas from von Neumann algebras to sphere packing bounds and other topics. The release of ten proofs with certificates and walkthroughs is presented as a significant milestone in the development of the model. This transparency allows the community to assess the quality of the work independently and to potentially extend the work in new directions. The mention of nonsofic groups existence is highlighted as one of the new results proved by the model, adding to the body of knowledge in group theory.

> yes, nonsofic groups exist: this statement is one of many new beautiful results proved by Astra, our next major model. We're releasing 10 such Astra proofs, complete with lean certificates and CoT walkthroughs for each of them. The results are wide-ranging, from von Neumann algebras (disproof of Connes' Rigidity Conjecture) to better bounds for high dimensional sphere packing, for circuit complexity, for monochromatic triangles in multicolored graphs, and more.Sébastien Bubeck, OpenAI researcher

Greg Brockman also commented on the achievement, noting the ten significant advances achieved using the internal Astra model at a total cost of about two thousand dollars. This cost figure provides a benchmark for the efficiency of the model in tackling these complex problems that have challenged human mathematicians. The combination of the announcement and the public repository supports the claims made about the model's performance in formal reasoning tasks. Experts outside the company are expected to review the proofs in the coming weeks to confirm the validity of the disproof and other results, which will determine the long-term impact of this work on the mathematical community.

## What comes next for OpenAI and frontier models in mathematics?

Looking ahead, OpenAI is likely to continue integrating such capabilities into its models, potentially leading to more frequent contributions to open mathematical problems that remain unsolved. The success with Astra suggests that scaling and refinement of reasoning abilities will yield further results in areas like quantum complexity and lattice cryptography as the technology matures. The community can expect additional releases or updates to the GitHub repository as new proofs are generated by improved versions of the model. This trajectory points to a future where AI plays a more integral role in the advancement of theoretical fields by providing new tools and insights.

The implications extend to education and training, where students and researchers might use similar tools to explore complex topics and verify their own conjectures. The formal nature of the proofs ensures that they can be used as reliable references in academic work and further research. As more models adopt this approach, the pace of mathematical discovery could increase, benefiting fields that rely on these theoretical foundations for practical applications. The current results set a precedent for what is possible with current frontier model technology in the domain of mathematics and computer science, encouraging further innovation in the area.

The potential for AI to contribute to mathematics is still being explored, but the current results provide a strong indication of what is possible with current technology. The combination of generative AI with formal proof systems like Lean creates a powerful tool for discovery. This could lead to breakthroughs in areas that have been stagnant for decades, as the model can explore possibilities that human researchers might overlook due to time constraints or cognitive limitations. The future development of these models will likely focus on improving the accuracy and scope of the problems they can solve, leading to even more impressive results in the coming years.

In conclusion, the work by the Astra model represents a significant milestone in the application of frontier AI to theoretical problems. The low cost and high impact combination makes it an attractive area for further investment and research. The community is encouraged to engage with the released materials to fully appreciate the scope of the achievements and to contribute to the ongoing development of AI in mathematics. This is just the beginning of what such models can accomplish in the field.

- Examine the released Lean certificates for each of the ten proofs on the GitHub repository.
- Review the chain of thought walkthroughs to understand the model's reasoning process.
- Verify the formal proofs independently using the Lean 4 theorem prover.
- Consider the implications of these results for related open problems in the respective fields.
- Monitor future announcements from OpenAI regarding additional advances by the Astra model.

## Sources

1. [We provide new results for the following problems. The results were achieved by an internal version of Astra, our next major model. ... Connes’s rigidity conjecture. Disproof of a longstanding conjecture that certain groups are uniquely determined by their von Neumann algebras](https://openai.com/index/ten-advances-in-mathematics/)
2. [This repository contains Lean 4 formalizations of the results presented in Ten advances in mathematics and theoretical computer science by OpenAI. ... Connes’s rigidity conjecture: A counterexample to the conjecture that certain groups are determined by their group von Neumann algebras. (ConnesRigidity.lean)](https://github.com/openai/ten-proofs)
3. [ten significant advances in mathematics and theoretical computer science. solved using an internal version of Astra, our next major model, for a total cost of about $2000 at Sol API prices](https://x.com/gdb/status/2083457463337287721)
4. [yes, nonsofic groups exist: this statement is one of many new beautiful results proved by Astra, our next major model. We're releasing 10 such Astra proofs, complete with lean certificates and CoT walkthroughs for each of them. The results are wide-ranging, from von Neumann algebras (disproof of Connes' Rigidity Conjecture) to better bounds for high dimensional sphere packing, for circuit complexity, for monochromatic triangles in multicolored graphs, and more.](https://x.com/SebastienBubeck/status/2083456300692979886)

---
Source: https://aiintelreport.com/frontier-models/openai-astra-ten-math-advances
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt