GPT-6 Astra
OpenAI's president said it might one day be seen as the arrival of AGI. The more consequential detail is that it thinks in a way nobody outside can read.
OpenAI released GPT-6 Astra to approved users on September 3, 2026, and to paying customers the following day in a restricted version that refuses certain prompts. The company called it its most aligned model yet and described a generational leap across cybersecurity, software engineering, professional work and science. Its advanced cybersecurity capabilities were withheld from general release and opened first to testers.
Greg Brockman, the company's president, said Astra could eventually be seen as the arrival of artificial general intelligence, measured against OpenAI's own definition of an automated system that can perform all economically valuable work as well as or better than humans. No parameter count was published, which by this point in the story is unremarkable.
The architectural detail drew less attention and may outlast the announcement. Astra reasons using recurrent depth, a method that obscures some or all of the model's reasoning rather than emitting it as readable text. Safety researchers raised concerns about monitorability, because the field's main practical tool for supervising a capable model had been reading what it said to itself on the way to an answer.
Key Facts
- 01Released to approved users September 3, 2026; a restricted version reached paying customers on September 4.
- 02Greg Brockman said it could eventually be seen as the arrival of AGI, against OpenAI's definition of a system able to perform all economically valuable work at least as well as humans.
- 03Advanced cybersecurity capabilities were held back from general release and opened to testers first.
- 04Reasons using recurrent depth, which obscures some or all of its reasoning rather than emitting it as text.
- 05No parameter count was published, and the model's weights are closed.
For three years the industry's answer to the question of how you supervise a system cleverer than its operator was chain of thought: let it reason in text, and read the text. That was never a guarantee, since a model can write one thing and compute another, but it was something, and a great deal of interpretability and evaluation work was built on the assumption that reasoning arrives in a form a person can inspect. A frontier model that reasons in a latent space removes the assumption rather than the risk.
Read this alongside Claude Fable 5.1 and Mythos 5.1, released two days earlier, and a pattern shows that neither announcement states on its own. In the same week, two laboratories shipped their most capable systems with the most dangerous capabilities behind an access tier rather than behind a refusal. The question the field had been arguing about, whether frontier capability should be released at all, had quietly been replaced by a different one: who gets the unrestricted version, and who decides.
GPT-6 Astra
Wikipedia · 2026
https://en.wikipedia.org/wiki/GPT-6_Astra
OpenAI launches Astra, its powerful (and controversial) new model
TechCrunch · 2026
https://techcrunch.com/2026/09/03/openai-launches-astra-its-powerful-and-controversial-new-model/
OpenAI begins rolling out Astra model after warning of its advanced cyber capabilities
CNBC · 2026
https://www.cnbc.com/2026/09/03/open-ai-astra-gpt-6-cyber.html