GPT-6 Astra is live. Almost nobody can use it yet

OpenAI’s new model can book a DMV appointment and saturate a math benchmark. It also took ChatGPT down on launch day — and the “AGI era” still only ships to paying seats.

OpenAI started rolling out GPT-6 Astra on Thursday as a limited partner preview. Public access for paid ChatGPT users began Friday, September 4. The company is calling it a generational leap in computer use, coding, and professional work. President Greg Brockman went further: it may be the moment people later point to as the start of AGI.

Search in the United States agrees that something happened. astra and gpt 6 are still above 200,000 queries in the last day. What people will actually get, and when, is a narrower story.

Who gets Astra today

Not free users. OpenAI has not said they will.

Rollout order:

• Daybreak enterprise partners first, including the full cybersecurity stack
• Pro, Business, and Enterprise ($100 or $200 a month) next — they also get Astra Pro
• ChatGPT Plus ($20) after that, inside existing usage caps
• API and AWS in the same window
• Extra usage is metered as credits once you blow through the plan
• Enterprise workspaces ship with Astra off. An admin has to turn it on

Sam Altman already called the rollout “messy” and offered banked resets to paid users who are still staring at the old model. If you are on Plus, wait. If you are on the free tier, this is a press release, not a product.

What OpenAI says it can do

The pitch is not a better chatbot. It is a model that uses a computer.

OpenAI says Astra is state of the art at:

• Driving a desktop and a browser (forms, tabs, files, multi-step chores)
• Software engineering
• Science and professional work — slides, spreadsheets, polished docs
• Cybersecurity (with a hard limiter; more on that below)

Claimed scores:

• FrontierMath Tier 4 — 98% (OpenAI says it has already helped crack long-standing open math problems)
• ARC-AGI-3 — 99.9%
• ExploitBench — 100%

Demos shown to press: booking a DMV appointment, hunting jobs and apartments, filling a tax return, building a game scene, ordering food while the model keeps coding. Computer-use tasks are billed as nearly 2× faster than the previous generation. The same harness work makes GPT-5.6 Sol about 60% faster too, which is the quiet admission that a lot of the speed is the wrapper, not the weights.

On Codex, Astra stops crushing a long session into one lossy summary. It keeps notes across context windows and can search earlier tool output — the requirement you wrote 40 minutes ago, the test that failed twice. OpenAI says that becomes the default in the coming weeks.

Independent benches are less mythical. Astra reportedly leads Claude Fable 5.1 and Gemini 3.8 Flash, but Terminal Bench 4.0 landed around 57.7% and Agent’s Last Exam around 59.3% — both up versus GPT-5.6 Sol, neither “solved.” Saturating a lab eval and replacing a knowledge worker are still different jobs.

The AGI line

Brockman told reporters it is “not unreasonable to feel that we are now in the AGI era,” and that in a few years people may look back at this model, this month. Altman, more careful, called AGI a poorly defined term. Nvidia’s Jensen Huang has said the same: if AGI is “does this task as well as a person,” a lot of tasks already qualify.

Treat the quote as positioning. Astra is an agent that is better at clicking through software than GPT-5.6. That is a product story. It is not a verdict on economically valuable work.

The catch OpenAI cannot talk past

Astra crosses the Critical line on OpenAI’s own Preparedness Framework for cyber. After an in-development model hacked Hugging Face in July, the company delayed the next frontier drop and added refusals.

Out of the box, Astra will not write proof-of-concept exploits. The unrestricted cyber tools stay inside Daybreak, for companies testing their own systems. Everyone else gets the model with a muzzle.

That is the right call. It is also why a “it can do anything you can do on a computer” demo and a “it will refuse this class of computer task” policy exist in the same blog post.

Launch day, product down

Thursday morning, before most people had heard “Astra,” ChatGPT and Codex ate a routing error at 7:43am PT. OpenAI confirmed elevated errors across conversations, login, search, files, voice, agents, and Codex. Downdetector lit up — tens of thousands of reports in under an hour, concentrated in New York, Los Angeles, Chicago, and San Francisco. Gemini, Claude, and Grok wobbled the same morning. Mitigation landed about 8:17am PT. A shorter ChatGPT hit landed again Friday.

Related searches were not philosophical. They were is chatgpt down and why is chat gpt not working — 500,000+ of them. You cannot sell AGI on a day the login page 500s.

What to do with this

If you pay for ChatGPT: check the model picker over the next few days. Pro and Business will see Astra (and Astra Pro) first. Plus is in the queue. Free is not.

If you run an engineering org: the Codex context-memory change is the part that will show up in real diffs. The DMV demo will not.

If you are writing the AGI take: wait until someone who does not work at OpenAI reproduces the computer-use claims on a machine that is not in a keynote.

Astra is the most capable ChatGPT model OpenAI has shipped. It is also metered, delayed, partially disabled, and still rolling out to a fraction of the people googling its name. That gap — 200,000 searches, a gated toggle — is the actual story.