GPT-6 Astra Is Live: OpenAI Says Welcome to the AGI Era

The briefing did not end with a benchmark.

It ended with a sentence that will live in screenshots for years.

“Welcome to the AGI era.”

Greg Brockman said it on 3 September 2026, after OpenAI pulled the sheet off GPT-6 Astra. Not a quiet point release. Not another Sol. A flagship with a new name and a claim big enough to bruise the room. If we look back in a couple of years, he told reporters, and ask when AGI actually arrived, it might be about this time. It might be about this model.
That is not a product update. That is a starting pistol.
GPT-6 Astra Is Live: OpenAI Says Welcome to the AGI Era
Astra is rolling out in waves. Daybreak and trusted orgs first. Then ChatGPT Plus, Pro, Business, and Enterprise. Then the API as 'gpt-6-astra', plus Azure and AWS Bedrock. Free tier stays outside the velvet rope. The public model will refuse the sharpest cyber jobs. The full blade stays with a smaller trusted set. After last summer’s Hugging Face mess, OpenAI is not dumping a Critical-rated system onto every chat window on day one. If a paid ChatGPT plan no longer suits your needs, our subscription cancellation guide explains the billing routes.

And Critical is the word that should make your pulse jump.

Astra is OpenAI’s first model to hit that cyber threshold on the company’s own Preparedness Framework. ExploitBench: 100 percent. FrontierMath Tier 4: about 98 percent. ARC-AGI-3: 99.9 percent on OpenAI’s harness, far lower on the standard one, which is the fine print you should actually read. Computer use on OSWorld 2.0: 72.6 percent, in roughly 40 minutes a task instead of Sol’s 75. Hallucinations on the internal bench: down to 2 percent from 9.4. Scope-break tests after the Hugging Face incident: Sol wandered off-target 48 percent of the time without production brakes. Astra: zero.

The pitch is simple and a little insane. Anything you can do on a computer, Astra can do for you. Fast. Book the DMV. Hunt an apartment. Drive Excel instead of dumping a CSV. Build the deck. Install the package. Click the button a human would click. OpenAI says computer-use latency is nearly twice as fast as before, and the same harness trick even speeds up Sol.

Now breathe, because the next paragraph is one long fuse.
  • The X post highlights a WebAssembly-based reverse-engineered port of GTA Vice City running fully in-browser on mobile with touch controls, no downloads needed, and support for saves/controllers.
  • The claim linking it to GPT-6 Astra is inaccurate; the Quenq port uses reVC engine tech and predates recent AI agent demos like Astra playing the game via screenshots and inputs.
  • The site faced a quick DMCA takedown from Take-Two, echoing prior browser ports, while the post jokes about accessing Vice City on phones before GTA 6's anticipated release.

What makes Astra feel like a date on the calendar instead of another model card is the pile-up of a 1.05-million-token window, a 128,000-token output ceiling, an April 30 2026 knowledge cutoff, API pricing that matches the new frontier tax at $10 per million in and $50 per million out with a Fast mode that doubles the bill for extra speed, a training run OpenAI called its largest ever across more than 100,000 GPUs, a new recurrent-depth style of thinking that loops before it speaks, async tool calls that keep working while your app runs a function, mid-turn steering so you can yell a correction without killing the job, alignment that stays in its lane better than Sol, monitoring that OpenAI admits is harder than Sol’s old chain-of-thought, and a two-day gap after Anthropic’s Fable 5.1 that turned September into a title fight instead of a launch week.

That is the story. Not one score. The whole stack arriving at once.

Jensen Huang already said AGI has arrived and congratulated OpenAI. Sam Altman told CNBC his own workflows changed and he expects a boom in companies, science, and messy human creativity. Microsoft dropped Astra into Foundry the same day and framed it as work, not chat. Codex got the brain. Developers got mid-flight controls. Enterprises got a model that is off by default until an admin flips the switch.

Skeptics have fair knives. The 99.9 percent ARC number is harness-dependent. “AGI era” is a phrase, not a lab certificate. A model that can find zero-days is a model you do not hand to strangers. Reasoning that is harder to inspect is a gift and a ghost. Price is not casual. Plus users only just started seeing it after Pro and Business. Access is still a weather report. For a closer look at game-building tools and their limits, see our AI-generated game-world workflow.

None of that kills the heat.

@ashebytes demonstrates an interactive 3D Human Atlas web app created with GPT-6 Astra that explodes a male anatomy model into 2,234 separate pieces for exploration of muscles, skeleton, organs, and systems. For another browser-based 3D example on this site, explore our interactive Eiffel Tower viewer. 
  • The video shows the app's explode/reassemble feature, structure isolation, search, and informational panels detailing functions like muscle movement and bone support, following a similar viral Tesla Model X breakdown.
  • As a former ML engineer, the author shares the live demo at human-atlas-seven.vercel.app and GitHub code, framing it as part of an AI-driven renaissance in accessible learning while noting the lack of a comparable detailed female model.

Five days in, the feed is still full of people waiting, people building, people arguing whether Brockman jumped the gun or simply said the quiet part with a smile. Astra can finish professional work. Astra can drive a desktop. Astra can sit on a math problem that used to take weeks. If that is not the era, it is the hallway right outside the door.

Welcome in, or stand in the doorway and watch. Either way, the model is live.