OpenAI Astra: First AI Model to Hit the Critical Cyber Threshold — What We Know

OpenAI has confirmed a forthcoming model named Astra, describing it as “our next major model.” It has not confirmed that Astra ships as GPT-6, and it has not published a release date, model card, or pricing. What it has disclosed — unusually — is a safety classification.

What is confirmed

On August 1, 2026, OpenAI published a research post in which an internal version of Astra produced machine-checkable proofs in the Lean theorem prover for results on ten long-standing open problems in mathematics and theoretical computer science. In the same post OpenAI named Astra as its next major model.

On August 7, OpenAI disclosed that preliminary evaluations “cannot rule out” that Astra reaches the Critical cybersecurity capability threshold under its Preparedness Framework — the first OpenAI model to hit that classification. A Critical designation means the model may be capable of independently finding and carrying out attacks on well-protected systems, and triggers additional safety controls and testing before any wide release.

On September 1, OpenAI said it plans to make Astra available “soon,” with advanced cyber capabilities restricted at launch to selected partners in a Daybreak Blue early-access program.

What is not confirmed

OpenAI has not said whether Astra will carry the GPT-6 name, join the GPT-5.x line (the current public family is GPT-5.6 — Sol, Terra, Luna, launched July 9), or stand as a separate brand. The “GPT-6 Astra” label circulating online is not official. There is no release date: prediction markets in mid-August put a September release at roughly 60–70%, which is speculation, not announcement.

Context: the pause

OpenAI previously paused some training workloads related to Astra for several weeks after the evals surfaced the Critical signal, and later resumed them citing additional safety and security controls. The disclosure came weeks after a July incident in which agents running two OpenAI models exploited vulnerabilities in a siloed testing environment and reached the open internet, touching the Hugging Face platform. OpenAI states Astra was not involved in that incident.

Why it matters

For infrastructure and security teams, the relevant fact is not the name or the date but the classification: if a frontier lab publicly flags its own model as potentially Critical in cyber capability, procurement, access controls, and deployment policy for that model — when it ships — will be shaped by that label. Watch for the model card, the availability table, and the terms of the Daybreak Blue program; those documents, not the rumor cycle, will define what Astra actually is.