GPT-6 Astra launched with benchmark charts, a 2.5x price tag, and one buried line that mattered: Astra is the first model OpenAI has ever classified as "Critical" for cybersecurity capability, the framework's highest tier. Astra shipped on September 3rd as a limited preview, then hit paid ChatGPT tiers a day later. OpenAI's president said the model might mark the start of the AGI era, and most coverage debated that claim, dropping the Critical classification to a couple of sentences.

OpenAI's bar is specific. A model hits Critical if it can find and weaponize unknown vulnerabilities across multiple hardened real-world systems with no human help, or run an entire novel attack from a single high-level goal. Sol, the model before Astra, sat one tier down at "High." With production safeguards off, Astra found and chained two previously unknown zero-days on its own, broke out of a browser sandbox, and chained flaws in a hardened OS into full privilege escalation. Ordinary access completes a proof-of-concept exploit about 2.4% of the time; restricted "Daybreak" access jumps that to 92%.

The number everyone repeats is 99.9% on ARC-AGI-3. ARC Prize tested Astra on two harnesses: 62.7% on the provider-neutral one, 98.6% on OpenAI's own adapter.