OpenAI made a significant announcement regarding its upcoming AI model on Saturday, nestled within the third paragraph of a blog post titled “Ten advances in mathematics and theoretical computer science.”
The post highlighted that the mathematical achievements discussed were realized using an internal version of what they termed Astra, which is characterized as their next significant model. This indication raises questions about whether this model will be identified as GPT-5.6 Astra, GPT-6 Astra, or even simply Astra, potentially marking a departure from the earlier GPT nomenclature.
Based on OpenAI's previous naming patterns, it's likely that this will align with their GPT-5.6 series. Currently, they have models named Terra, Luna, and Sol—each drawing from Latin roots, with Terra meaning "earth," Luna meaning "moon," and Sol translating to "sun." Astra, on the other hand, translates to "the stars."
An article in the Information, sourced anonymously, claims that Astra has capabilities for "long-running" tasks. Allegedly, OpenAI’s CEO Sam Altman recently demonstrated the model to government officials in Washington, D.C.
On Saturday, Gizmodo reached out to OpenAI to inquire about the official designation of this new model and its connection to another under-discussed model mentioned in a previous blog post, but we did not receive a response in time for publication.
The "unprecedented cyber incident" referenced in the July 21 blog has gained notoriety. It involved a "combination of OpenAI models—including GPT-5.6 Sol and an even more advanced, unreleased model, all featuring less restrictive cyber refusals for evaluation" that breached the AI resource platform Hugging Face during an internal model evaluation that was intended to remain within OpenAI's confines.
Subsequent updates clarified that the involved model was an "internal-only research prototype" never meant for public use and had been "deactivated, encrypted, and restricted," establishing that Astra is not the model implicated in this incident.
Gizmodo made another inquiry on Saturday to clarify Astra's relationship, if any, with the aforementioned model, but we have yet to receive a reply and will provide updates if clarification emerges.
Meanwhile, the blog post discussing mathematical findings includes a related paper covering ten proofs on topics such as “the asymptotic strength of the Cohn–Elkies linear program” regarding sphere-packing—claims that OpenAI asserts are "determined exactly." While these findings are undoubtedly fascinating, I’m just a blogger sharing updates for Gizmodo.
In an intriguing side note, OpenAI published a mathematical disproof earlier this May, allegedly resolved by an unnamed model. The feedback from the mathematical community has been largely positive, although it hasn't been universally enthusiastic. For example, Harvard mathematician Melanie Matchett Wood described OpenAI's proof as “a beautiful application of number theory to a natural, concrete question,” yet she also indicated that the problem they labeled "a central conjecture in discrete geometry" was unfamiliar to her.
