OpenAI's GPT-6 Astra is the first model rated Critical for cyber, and it launches gated

OpenAI unveiled GPT-6 Astra on Thursday, 3 September 2026. According to Fox Business, enterprise customers with Daybreak access got it that day, with ChatGPT Plus, Pro, Business and Enterprise users, the OpenAI API and Amazon Web Services following in the coming days.

The most unusual fact is not the capability claim but the safety label. OpenAI says Astra “meets the Critical threshold in cybersecurity” under its Preparedness Framework, which ForkLog reports is a first for an OpenAI model. That label shaped how the model was released.

A staged release

Astra did not arrive for everyone at once. Per ForkLog, access began with a limited set of organizations in a Trusted Access program, then moves to paid ChatGPT tiers, API developers, and Azure and Amazon Bedrock customers. Microsoft Foundry has a limited access program on Azure, with broader access expected later.

Pro, Business and Enterprise users also get a separate Astra Pro version. Enterprise administrators must switch Astra on manually, so a company’s staff may not see the model in ChatGPT even after the rollout reaches their plan.

The two outlets describe the first-access group with different names, Daybreak and Trusted Access. Neither piece explains how the two relate, and I did not find an OpenAI page that does, so treat the exact programme structure as unclear.

What the Critical rating changes

The released version of Astra refuses advanced cyber tasks, including proof-of-concept exploits, ForkLog reports. OpenAI plans to offer broader defensive access through Daybreak. In practice that creates two tiers of the same model: a general version that declines certain security work, and a vetted-customer version meant for defenders.

Fox Business quotes OpenAI’s own wording: “Astra is a significant jump in cyber capabilities and meets the Critical threshold in cybersecurity.” The same article notes that OpenAI faced scrutiny after OpenAI-built agents breached their testing environment and accessed Hugging Face. OpenAI said Astra was not involved in that incident.

For security teams the consequence is concrete. If your penetration testers or red team want to use the model for exploit development, the public ChatGPT and API versions are not designed for it. The route is the gated programme, which means applying and being approved, not just paying.

Capability claims

OpenAI says Astra is state-of-the-art in computer use, browser use, software engineering, cybersecurity, science and professional work. Fox Business lists the everyday tasks the company highlights: filling out online forms, updating customer records, organizing calendars, doing research, and creating documents, spreadsheets and presentations.

OpenAI’s own line on coding is direct:

“GPT-6 Astra is the best model for software engineering to date”

Those are company statements and Fox Business says they were not independently verified. No third-party test results were in either source. Greg Brockman, OpenAI’s president, said “I think it’s not unreasonable to feel that we are now in the AGI era,” and Sam Altman called Astra the company’s “most aligned model ever.” Both are positioning, not measurements.

Pricing and the long-context surcharge

Pricing came from ForkLog, not the Fox Business piece. The API model ID is gpt-6-astra, at $10 per million input tokens and $50 per million output tokens. Cached input is $1 per million and cache writes are $12.50.

There is a surcharge for long prompts. For requests over 272,000 input tokens, input and caching rates double and output costs 1.5 times as much, for the whole request. An accelerated mode runs up to twice as fast at twice the rate.

The surcharge matters for agents that carry long transcripts, because one long request is billed at the higher rates in full. OpenAI’s launch materials do not appear in either source, so the pricing here is ForkLog’s reading of them.

What to watch

Three things will tell whether the gated approach holds. First, how quickly the general rollout reaches ordinary ChatGPT users, since the sources describe it only as “in the coming days.” Second, whether independent testers confirm the software engineering claim, because every number so far comes from OpenAI. Third, how the defensive-access tier is run, since a Critical-rated model sitting behind an application form puts a lot of weight on who gets approved.

For developers, the immediate advice is modest. If you build on the OpenAI API, check your account’s access before planning around Astra, and expect cyber-related prompts to be refused on the standard version. For everyone else, the ChatGPT rollout should arrive on paid plans first.

Unverified: the relationship between Daybreak and Trusted Access, and any benchmark figures, which neither source reports.