OpenAI released GPT-6 Astra on September 3, 2026. The company calls it the most intelligent and aligned model it has built so far, and its president Greg Brockman ended the launch briefing with a simple line: "Welcome to the AGI era." That single sentence is why this model is dominating tech searches right now.
What Astra Actually Does
Astra is not built around chat. OpenAI positioned it as a computer use and software engineering model. That means it can work inside browsers, spreadsheets, and desktop apps the way a person would, complete multi step tasks on its own, and produce finished documents or presentations instead of just explaining how to make them.
On benchmarks published by OpenAI, Astra scored 97.6% on FrontierMath Tier 4, compared to 83% for the previous model, GPT-5.6 Sol. It also scored 72.6% on OSWorld 2.0, a test for computer use tasks, up from 65.7%. These numbers come from OpenAI itself and have not been independently verified yet.
The Cyber Security Delay
Astra was supposed to launch about four weeks earlier. OpenAI held it back after internal testing showed it could find and chain security vulnerabilities on its own, including two real zero day flaws that OpenAI reported to the affected software makers. This pushed the model past what OpenAI calls its "Critical" cybersecurity threshold.
Because of this, the public version of Astra is trained to refuse tasks like writing working exploit code. The full, less restricted capability is only available to vetted organizations through a program called Daybreak.
Rollout and Pricing
The rollout is happening in stages. Select organizations got access first on September 3. ChatGPT Plus, Pro, Business, and Enterprise users are getting it over the following days, along with access through the OpenAI API and Amazon Web Services.
On the API, Astra costs 10 dollars per million input tokens and 50 dollars per million output tokens. That is two and a half times the price of GPT-5.6 Sol. It comes with a 1,050,000 token context window and a knowledge cutoff of April 30, 2026.
Is It Really AGI
Brockman said he personally believes OpenAI has reached AGI, but he left the final judgment open to users. Not everyone agrees. Toby Walsh, an AI researcher at the University of New South Wales, said that current AI intelligence is still "jagged," meaning these models can do very advanced things while still failing at simple tasks. So the AGI claim is a mix of an official company statement and an ongoing debate, not a settled fact.
What This Means Going Forward
The bigger pattern here is not just one model release. Meta, Google, and Anthropic have all pushed out major model updates in the same week, and government scrutiny over how these systems are released is increasing. OpenAI's launch also passed through a level of government review, which suggests future frontier model releases will be shaped as much by policy as by engineering readiness.