OpenAI, the developer of ChatGPT, has disclosed part of the performance of its next-generation artificial intelligence (AI) model "Astra," which is undergoing internal testing.
OpenAI said on the 2nd (local time) that, using Astra, it solved 10 math problems that had remained unsolved for at least 10 years. The problems OpenAI introduced include high-dimensional geometry, group theory, a branch of algebra, and quantum complexity.
OpenAI standardized and released Astra's proof processes in the form of Lean, a mathematical proof verification tool.
OpenAI said the total token expense Astra used to solve the problems would be about $2,000 (about 2.9 million won) when calculated based on the application programming interface (API) pricing of "GPT-5.6 Sol."
According to The Information, a U.S. information technology (IT) outlet, OpenAI CEO Sam Altman recently gave an in-person demo of Astra to political figures and regulators in Washington, D.C.
Altman was said to have explained that Astra incorporates a capability in which multiple AI agents collaborate for long periods to solve difficult problems.
Meanwhile, when OpenAI released GPT-5.6 recently, it assigned celestial-themed names by performance tier such as "Sol" (sun), "Terra" (earth), and "Luna" (moon). Given that, Astra, which means star, is presumed to be a similar submodel name of GPT.
However, it was reported that OpenAI has not yet decided whether to release Astra as an additional model in the 5 series, such as GPT-5.6 or 5.7, or to debut it as GPT-6. Sources said OpenAI plans to submit the model's performance and safety to the government for review before Astra's release.