OpenAI releases hundreds of mathematical results from an internal model not yet launched

OpenAI announced, on October 6, 2026, the publication of a large set of mathematical results obtained by an internal model that has not yet been commercialized. The material was placed in a public GitHub repository, accompanied by documentation that includes formalizations in Lean, reasoning summaries, and computation statistics. The company stated that this decision followed the recommendations of the independent advisory group on Mathematics and AI at the Institute for Advanced Study, even though it chose not to disclose the model’s source code or all the prompts used in the experiments. This announcement was presented as an important step toward transparency and collaboration between artificial intelligence and mathematical research, while emphasizing safeguards to protect the model’s intellectual property.

Context and motivation

In recent years, OpenAI has heavily invested in systems capable of generating formal proofs and pushing the frontiers of mathematical research. The success of earlier versions, such as GPT‑4‑Turbo, created expectations that AI could accelerate the resolution of open conjectures and reduce researchers’ workload. In this scenario, the independent advisory council of the Institute for Advanced Study recommended that the company publicly share the results obtained, to enable external evaluation and foster collaboration between AI and the academic community. This recommendation aimed to encourage independent validation and stimulate collaborative innovation.

The decision to publish only the results, and not the underlying model, reflects a risk‑mitigation strategy that OpenAI has adopted in other sensitive areas. By making the proofs and reasoning summaries available, the company seeks to provide enough material for specialists to verify the validity of the results, while protecting the model’s intellectual property and avoiding potential misuse before a deeper safety assessment. This balanced approach aims to reconcile scientific openness with ethical responsibility.

What was released

The released package contains several categories of information. Each result includes a formalization in Lean – the computer‑assisted proof language – and a summary of up to ten steps describing the model’s reasoning. Additionally, OpenAI supplied statistics on the number of problems attempted, the success rate, and a computational cost estimate: the average result consumed roughly three hours of ChatGPT Pro reasoning time. Not all prompts that generated the demonstrations were included, nor was the source code of the model that produced the proofs. This selection of information aims to give a comprehensive overview while preserving sensitive components.

• Complete Lean formalizations for 372 results
• Ten reasoning summaries per result
• Attempt statistics and success rate
• Computational consumption estimate (≈ 3 hours of ChatGPT Pro per result)

Scientific American counted 372 results that, according to the release, solve or advance open questions across various fields of mathematics, from number theory to algebraic topology. The magazine highlighted that, although several proofs have been formally verified in Lean, the community still needs to assess the importance and originality of each contribution. OpenAI acknowledged that formal verification alone does not guarantee scientific novelty and that a peer‑review process will be essential to validate the real impact of the findings. This acknowledgment shows an awareness of the limits of automated validation.

Community reactions

Mathematicians who examined the repository expressed skepticism about the complete verifiability of the results. Some noted that, without access to the model or the full prompts, it is difficult to reproduce the reasoning path that led to the demonstrations. Others questioned the novelty of the proofs, suggesting that some may correspond to already known results or trivial variations of existing theorems. The review workload was also cited as an obstacle: evaluating hundreds of formal proofs requires time and resources that many academic institutions do not have.

Despite these reservations, part of the community welcomed the initiative as a significant step toward integrating AI and mathematics. The public availability of Lean formalizations allows researchers to test extensions, explore new conjectures, and develop reasoning‑support tools. Some experts emphasized that publishing results, even without the model, creates a starting point for discussions on evaluation standards, credit attribution, and ethics in AI‑assisted knowledge production. This perspective opens the door to constructive debate.

Next steps and implications

OpenAI pledged to improve citation quality, exposition clarity, and overall presentation of the results in future iterations. The company also announced funding for workshops that will bring together AI researchers and mathematicians, with the goal of defining best practices for verification and publication of machine‑generated results. Meanwhile, the potential benefits – accelerated discoveries, democratization of proof tools, and new interdisciplinary collaborations – must be weighed against current limitations, notably the lack of full model transparency and the need for rigorous human review. What changes now is that the community has access to a large corpus of formal proofs that can serve as a foundation for future investigations, while a critical debate unfolds about how to validate and integrate AI contributions into the traditional scientific landscape.