OpenAI has officially confirmed the creation of a new artificial neural network named Astra. According to the published report, the model represents the company's "next major family" of developments and has demonstrated the ability to solve fundamental scientific problems that humanity has struggled with for decades.
Breakthrough in Mathematics and Group Theory
The internal version of Astra successfully solved ten open problems in mathematics and theoretical computer science. Researchers had worked on these tasks unsuccessfully for at least ten years, and in some cases, significantly longer. The model's areas of coverage include multidimensional geometry, coding theory, quantum complexity, lattice-based cryptography, and extremal combinatorics.
One of the most significant achievements was the construction of the first example of a non-sofic group. The existence of such a mathematical object has been discussed for many years, and its discovery could be the key to solving one of the fundamental problems of modern group theory.
Reaction of the Scientific Community
The results of Astra's work have sparked a lively response in the academic community. Thomas Bloom, a mathematician from the University of Manchester, called them "big news." In his opinion, the significance of these discoveries exceeds even the counterexample to the unit distance hypothesis published in May.
At the same time, Bloom is skeptical about the idea of artificial intelligence soon replacing mathematicians. He notes that modern AI systems are themselves based on the decades-long work of the mathematical community and are trained on scientific literature created by them.
Technology and Formalization of Proofs
Noam Brown, one of the developers of Astra's reasoning technology, stated that the company has already tried to apply the model to other well-known open problems, but so far without success. In particular, the model has not yet been able to solve any of the "Millennium Prize Problems" for which the Clay Mathematics Institute offers a million dollars.
Nevertheless, the volume of computation required to solve the ten current tasks was relatively small: according to OpenAI estimates, this is equivalent to approximately $2,000 using the Sol model API at current rates. After generating ideas, researchers, together with Astra, turned them into full-fledged scientific papers.
An important aspect of the work was the use of the Lean system—a programming language and interactive tool for theorem proving. This allowed proofs to be formalized and made machine-verifiable, eliminating errors characteristic of manual notation. OpenAI also published a step-by-step description of the model's reasoning logic.
The Future of Astra and Regulatory Oversight
Previously, OpenAI reported on the development of a new family of models for performing long and complex tasks. This week, CEO Sam Altman presented Astra to representatives of the US government and regulatory bodies, emphasizing its ability to coordinate the work of multiple AI agents.
According to sources, Astra will enter the product lineup alongside the Sol, Terra, and Luna models. The commercial name has not yet been disclosed. The new models are already undergoing internal testing and will be the company's first systems to be evaluated under the advanced AI model assessment procedure being developed in the US.
The main task for developers remains ensuring the stable operation of models when performing long chains of reasoning, which is currently considered one of the main limitations of agentic AI systems.
