- Introduction to achievements in mathematics
- Details of complex mathematical problems
- Automated review and formal verification
- Expert criticism and lack of transparency
- Next steps and safety challenges
- The difference between problem-solving and theory-building
- Frequently asked questions
Introduction to achievements in mathematics
In early August, leading technology and research company OpenAI announced with great confidence—sparking widespread buzz and intense interest—that its new family of artificial intelligence models named “Astra,” which has not yet been officially released for public or commercial use, successfully generated innovative solutions to 10 highly complex open mathematical problems that had long baffled scientists. These solutions spanned highly intricate and abstract areas in pure mathematics and advanced theoretical computer science. To prove its credibility in this scientific achievement, the company released a detailed, fully documented 249-page scientific preprint accompanied by formal mathematical proof certificates that were fully automatically verified using the advanced “Lean 4” proof assistant. All of these documents were published for the public via repositories on the popular developer platform GitHub for review and evaluation. In a move highlighting the exceptional economic efficiency and computational power of this scientific breakthrough, the company stated in its explanatory statement that the total computing cost required to achieve these remarkable results was only about $2,000 USD, based on API pricing for the company’s advanced, high-performance “GPT-5.6-Sol” model.
Details of complex mathematical problems
According to the company’s official blog post, the operations and processing of the “Astra” model yielded unprecedented advancements and genuine discoveries in highly specialized and geometrically complex fields. These achievements included establishing new bounds in high-dimensional sphere packing theory, significant discoveries in advanced error-correcting code and mathematical theory, and the construction of the first explicit, well-defined non-sofic group—a complex and puzzling mathematical construction that has remained a riddle and sparked open questions among mathematicians and researchers since 1999 without a definitive answer or sound proof. In addition to this unique achievement, the model exceptionally proved the quantum parallel repetition theorem for general entangled games in quantum environments, delivering decisive results regarding the complex Erhard volume conjecture and the stiffness conjecture by mathematician Kunz. Furthermore, based on a detailed report published by the news outlet Quartz, the advanced system also intelligently solved and unpacked three additional complex problems derived from the famous and extensive catalog of prominent mathematician Paul Erdős, reflecting the system’s tremendous analytical capabilities in handling problems from multiple angles.
Automated review and formal verification
The company clearly stated in its explanatory note that its AI system independently and automatically undertook the task of formulating and generating the core, deep mathematical arguments. Following this, a team of specialized human researchers carried out the formalization and manual auditing of each mathematical proof produced by the model, using the advanced automated and programmatic proof assistant to verify the absence of logical fallacies. Of particular and distinctive interest, the error or unverified step counter within the published GitHub repository currently sits at absolute zero, serving as definitive technical proof that every logical step and mathematical sequence across all 10 proofs presented was mathematically checked and fully verified by the strict machine without any notable logical gaps that might question the validity of the final result or undermine its scientific credibility.
Expert criticism and lack of transparency
Despite this rigorous formal verification and intensive automated auditing of the mathematical proofs, many critics and academic experts quickly and seriously questioned whether these impressive, announced results truly represent a revolutionary leap and a watershed moment in the true cognitive capabilities of artificial intelligence, or merely data manipulation. In this context, prominent independent AI researcher Gary Marcus wrote a scathing critical article noting that a professional mathematician working for rival company Anthropic, named Levent Alperogullari, successfully and easily reproduced half of the results and proofs the company boasted about within just 24 hours, relying on the “Haiku” model—an already publicly available model rather than one with confidential capabilities. Based on these practical realities, Marcus strongly argued that this rapid success in replication raises serious, deep questions about whether the “Astra” model itself is the true, intelligent breakthrough, or whether the actual and significant progress lies simply in the ability to identify and select specific mathematical problems that respond positively and quickly to the automated search and verification techniques prevalent in current models. Criticizing the company’s methodology and lack of transparency, Marcus wrote: “They disclose absolutely nothing about the nature or number of complex mathematical problems they attempted to solve and failed miserably at. They simply gave us the numerator without revealing the denominator so we can see the true success rate, which is always a red flag in research and a cause for methodological skepticism regarding the credibility of the results.” He also drew attention to the fact that the company appears hesitant so far in deciding whether to name this model “GPT-6” as a revolutionary new generation, or simply “GPT-5.7,” hinting that this purported technical leap may be considered internally within the company’s halls as merely an incremental upgrade rather than the revolutionary breakthrough being rumored. Additionally, an ACM blog post described the achievement as remarkable yet vastly exaggerated purely for marketing purposes. Critics also did not fail to recall the “Leiden Declaration on AI in Mathematics,” signed by hundreds of prominent mathematicians in June 2026, where they openly expressed deep concerns regarding over-reliance on AI-generated complex proofs that may lack reliability and accurate attribution of intellectual credit for the achievement.
Next steps and safety challenges
To date, the model’s developer has not announced a specific or expected date for officially and openly releasing the “Astra” model to end users and commercial customers. This highly complex model will first need to pass rigorous reviews against federal AI safety standards and specifications to ensure it is not misused before seeing the light of day and reaching the hands of researchers and developers. For its part, the company explicitly acknowledged this ongoing controversy and escalating criticism surrounding the nature of its scientific achievement, issuing a statement affirming that it holds “deep respect and full understanding for those who feel concerned about the potential impact of this complex technology on the nature and reliability of scientific research,” and promising to publish full details and documentation of its adopted scientific methodology soon to calm fears. In a move aimed at winning over the academic community, the company recently launched a special, free edition named “ChatGPT for Academic Researchers,” providing free and unrestricted access for approximately 100,000 distinguished scientists and researchers to use its latest advanced models in their scientific research and experiments.
The difference between problem-solving and theory-building
Amidst this ongoing and provocative scientific debate, prominent and renowned mathematician Terence Tao delivered a valuable lecture on July 26—referenced by Marcus in his scathing critique—drawing a sharp line and establishing a fundamental, crucial distinction between two entirely different matters: first, the technical capability to find solutions to individual, isolated, and pre-defined mathematical problems—an area where artificial intelligence currently appears to achieve strong and promising results at first glance thanks to processing speed; and second, the complex, deep, and creative cognitive capability required to build and engineer an entirely comprehensive, coherent, and innovative mathematical theory from scratch to establish new fields of science. This latter domain remains purely human, with no definitive scientific evidence or tangible proof to date indicating that current artificial intelligence can approach it or achieve any noteworthy breakthrough in it.
Frequently asked questions
Question: What is the nature of the achievement claimed by the company regarding its new Astra model in mathematics?
Answer: The company announced that its intelligent model, Astra, succeeded thanks to its superior capabilities in solving 10 highly complex mathematical problems that had eluded solution and research for many years—such as constructing the first non-sofic group—at a surprisingly low computing cost that did not exceed $2,000.
Question: Why do experts and critics doubt the significance of this technical and mathematical achievement accomplished by the model?
Answer: Experts believe the achievement may be significantly exaggerated for promotional purposes, especially after an independent researcher managed to replicate half of these results with immense speed using other technical models available to the public. This suggests that the actual breakthrough lies in selecting problems that the machine can solve rather than the model’s own capacity for complex mathematical innovation and independent thinking.
Question: What is the difference explained by scientists between machine capabilities and human innovation in the field of mathematics?
Answer: Leading mathematicians clarify that there is a vast difference between a machine’s ability to reach a solution for a specific and pre-isolated problem through iterative search, and the brilliant mental capacity required to build comprehensive and coherent mathematical theories from scratch—an innovative task that remains exclusively the domain of the creative human mind to this day.
