Did OpenAI say GPT-6 Astra is AGI?

Not flatly. At the launch briefing OpenAI president Greg Brockman said that if you want to call it the first one, he thinks that is reasonable, and closed with 'Welcome to the AGI era.' The launch post itself leans on the 99.9 percent ARC-AGI-3 score.

What is the difference between 99.9 and 62.7 on ARC-AGI-3?

The harness. ARC Prize measured 62.7 percent with its standard harness and 99.9 percent with OpenAI's provider adapter harness. The model is the same; the tooling and memory around it differ.

What does the Critical cyber capability mean?

Under OpenAI's Preparedness Framework, Astra is the first model rated Critical for cybersecurity. OpenAI says that, with tools and access, it can find previously unknown flaws and develop exploits without step-by-step guidance, and that it found two real zero-days in internal testing.

What does Astra cost?

In the API, 10 dollars per million input tokens and 50 dollars per million output tokens at launch. ChatGPT Plus users get it only in some surfaces, and the rollout was described by OpenAI's CEO as messy.

OpenAI vs. the Truth: What Is Really Happening With GPT-6 Astra?

The AGI claim

OpenAI released GPT-6 Astra on September 3, 2026. At the briefing, president Greg Brockman said that if you want to call it the first AGI, “I think it’s reasonable,” and closed with “Welcome to the AGI era.” The launch post backs that with one number: Astra “saturates ARC-AGI-3 with a 99.9% score.” This video takes that number apart, then looks at the claims that matter more.

Watch the video (6:19): OpenAI vs. the Truth: What Is Really Happening With GPT-6 Astra?

What Astra does

The pitch is computer use. OpenAI says it is the company’s biggest training run, the first pretrained on more than 100,000 GPUs at its Stargate site in Texas. On OSWorld 2.0, which drops a model onto a real desktop and makes it do office tasks with a mouse and keyboard, Astra scored about 73 percent against about 66 for GPT-5.6 Sol, and finished tasks in around 40 minutes instead of 75. The demos are striking: a circuit board laid out in KiCad, a house modeled in Blender and turned into a walkable Unreal Engine scene, and an OpenAI employee having it rebuild San Francisco’s Palace of Fine Arts overnight.

Where 99.9 comes from

Read the footnote. Astra’s ARC-AGI-3 run used OpenAI’s own harness. The ARC Prize Foundation ran the model itself on its standard harness and got 62.7 percent, and 99.9 only with OpenAI’s provider adapter harness. A harness is the software around the model: which tools it can reach and what it remembers between steps. ARC Prize said maxing out the test “would not represent proof of achieving AGI,” and called Astra real progress, not AGI.

The independent index

Artificial Analysis runs an independent intelligence index, and OpenAI printed it on its own launch page: Astra 61.2, the outgoing Sol 60.9, and Anthropic’s Claude Fable 5.1 ahead at 65.7. Four days later the index moved to version 4.3 with harder coding and business-workflow tests. Astra scored 53, Sol dropped to 47, and Fable 5.1 landed at 53 as well. A real step up from Sol, and a tie with its biggest competitor. On the coding agent index Astra scored 67 to Sol’s 65. A gain, not a leap.

The cyber threshold

The claim we think matters most sits in the system card. Astra is the first model to reach the Critical level for cybersecurity under OpenAI’s Preparedness Framework. In OpenAI’s words, with the right tools and access it “can find previously unknown security flaws and develop new ways to exploit them” without a person guiding each step. In testing with safety limits off, OpenAI says it found and used two zero-days nobody knew about, both now reported to the maintainers. The public version refuses the most dangerous work, such as writing working exploits.

Verdict

Did OpenAI build AGI? It depends which number you pick, and that is the real problem: nobody agrees on what the word means. What changed this week for most people is nothing. Astra is still rolling out, Plus users get it only in some apps, and the API costs 10 dollars per million tokens in and 50 out. The number to watch is the office-work score. A model that can sit at a desk for 40 minutes and get the job mostly right is about to be sold to your employer. That is not AGI. It is automation with better manners.

Sources: OpenAI’s launch post and system card, ARC Prize’s results post, Fortune’s briefing report, and Artificial Analysis on benchmarking Astra and index v4.3.