OpenAI's new model, Astra, makes its debut: Tackles 10 unsolved mathematical problems in one go

According to DynaBeat monitoring, OpenAI has officially unveiled its next flagship model, Astra. The internal version has achieved new results on 10 long-standing unsolved mathematical and theoretical computer science problems. The core conclusions of these problems have seen no progress for at least a decade, with most being stagnant for even longer.

Some of these problems have been directly solved or overturned. Astra has constructed a non-sofic group for the first time, answering a key open problem in group theory; it has also overturned Connes' rigidity conjecture and resolved 3 Erdős problems. The remaining achievements include new upper and lower bounds on sphere packing, coding theory, quantum complexity, and post-quantum cryptographic problems with hardness proofs.

OpenAI stated that the model found the Tokens used to reach these 10 results, equivalent to approximately $2000 based on the Sol API price.

These mathematical proofs were generated by Astra. Humans then assisted in organizing them into papers, after which the model converted each proof into a Lean certificate for gradual computer verification of the validity of the derivations. OpenAI has publicly released the papers, proof certificates, and the model's reasoning process.