Ethan BrooksVIEW PROFILE →
OpenAI Astra Solves Ten Unsolved Math Problems for About 2,000 Dollars, Then Trips the First Critical Safety Flag in the Company History
OpenAI confirmed the name of a new model, Astra, that published verified proofs for ten previously unsolved problems, even as the company disclosed it could not rule out that the system reached a critical cybersecurity capability level.
OpenAI has confirmed the name of a new artificial intelligence model, Astra, doing so on August 1, 2026, inside a research post that focused on mathematics rather than a traditional product launch with a price and a release date.
The headline claim is striking, because the model was described as solving ten previously unsolved problems in mathematics and computer science, and publishing fully verified proofs on a public code repository at a computing cost of roughly two thousand dollars.
A new class of model

According to the company, Astra is not an upgrade to the existing GPT-5.6 system, but instead sits as a new class alongside the three tiers that OpenAI shipped in July, which it named Sol, Terra and Luna.
That framing matters, because it suggests the company sees Astra as a research grade system aimed at pushing the frontier of what these models can prove and discover, rather than a general assistant meant for everyday consumer use.
For the moment there is no release date, no model card, no public interface, and no price attached to the name, which leaves the wider world with a remarkable result but very little ability to test it directly.
The safety complication
On August 7, 2026, the company disclosed that it cannot rule out that Astra reached what it calls the critical cybersecurity capability threshold under its internal Preparedness Framework, the first time an OpenAI model has been placed in that category.
A critical designation is significant, because under the company own rules it triggers additional safety controls and testing that must be completed before any system carrying that classification can be released widely to the public.
That single disclosure helps explain why a model capable of producing verified mathematical proofs has still not been handed to users, since the same capability that impresses researchers can also raise concern among security teams.
The bigger picture
The announcement lands in a month of milestones for the company, which also saw its flagship chatbot cross one billion active users, a scale that turns even small changes in behavior into events felt by an enormous audience.
Taken together, the two developments capture the tension at the center of the field, where the same systems that can accelerate discovery are also the ones that demand the most careful handling before they are let loose.
For now Astra stands as a proof point rather than a product, a signal of how far the research has come, and a reminder that the gap between a laboratory result and a released tool is now defined as much by safety review as by engineering.






