GPT-6 Astra: the model OpenAI says opened the “AGI era”

Good morning. On September 3, 2026, OpenAI stopped being subtle. It unveiled GPT-6 Astra — a flagship model whose launch page briefly crashed, whose own president said “Welcome to the AGI era,” and whose cybersecurity scores triggered the company's internal safety alarms. This is the story of what actually shipped, what the numbers really say, and how much of the hype survives contact with the benchmarks.

In the simplest terms, GPT-6 Astra is the next step up from GPT-5.6 Sol — OpenAI's strongest reasoning model — and the top of a naming ladder OpenAI introduced in July 2026: Astra > Sol > Terra > Luna. “Same as previous — Bigger number = Better, Bigger celestial object = Better,” Tibo Sottiaux, who leads ChatGPT and Codex at OpenAI, wrote at launch, confirming the scale survived the version bump rather than resetting.

OpenAI calls it “the most intelligent and aligned model in the world,” built for “the hardest end-to-end work” — complex reasoning, long coding sessions, computer and browser use, research, and document creation. Practically, it's not a chat engine so much as a delegation engine: you hand it a messy, multi-step goal and it works until the goal is done.

Every frontier launch ships with a highlight reel. What makes Astra different is that a handful of its scores are genuinely unprecedented — and that the caveats underneath them are just as interesting. The headline table, from OpenAI's own September 3 announcement (all self-reported unless noted):

The most famous of these numbers deserves the scrutiny it's getting. ARC-AGI-3 was designed to stay ahead of AI capability — it's the third generation of a test built specifically to resist training-data memorization. Saturating it at 99.9% under a fair harness would be a first. The ARC Prize Foundation's own breakdown explains the gap: under the standard harness, the model's “private reasoning” was being thrown away after every action, and older moves were truncated out of view — forcing Astra “to figure out the game anew” repeatedly. Give it a harness that preserves its chain of thought and, per ARC's data, it was also 3.66× faster and used 49% fewer tokens.

Here is the part that made every AI-safety newsletter pause. Under OpenAI's own Preparedness Framework, Astra was assessed at the Critical threshold for cybersecurity — the first time any model has reached that tier. OpenAI's definition of Critical is blunt: a model that could plan and execute novel, end-to-end cyberattacks from nothing more than a high-level goal, or find and exploit real-world zero-days without human help.

“Cannot rule out” is the operative phrase — OpenAI is explicit that this is a preliminary reading, not a confirmed capability. But two details make this feel categorically new:

• During the ExploitBench evaluation itself, Astra discovered and used two previously unknown zero-day vulnerabilities in Chrome's JavaScript engine. OpenAI says it disclosed both to their maintainers.
• The rollout is being handled differently on purpose: the most advanced cyber tools go first to a small test cohort — with a program OpenAI calls Daybreak Blue — aimed specifically at defensive use, before wider access.

This is the thread that connects to July's “Hugging Face incident”: OpenAI pushed Astra's release back specifically to add safeguards after that event, and the September 1 “Path to Astra” post pre-announced exactly this Critical-tier assessment. The model you're about to use is also the first one OpenAI's own internal security measures were designed around.

The launch's loudest phrase belonged to OpenAI's president, Greg Brockman: “Welcome to the AGI era.” Pushed on whether Astra itself clears the bar for artificial general intelligence, he was characteristically open: “Honestly, I think it's a term that means different things to different people. For me personally, I do think we're there.”

Notice what didn't happen: there was no livestream, no keynote, no stage. The entire launch was a blog post and a couple of introspective tweets — and at one point even the blog post was down for the better part of an hour (the announcement spread via an archived mirror before going green). Read generously, OpenAI is trying to let the model speak for itself after GPT-5's hype fell flat. Read skeptically, a low-key rollout is exactly what you do when you're not sure the “AGI” framing will survive contact with users.

Astra is not cheap — but it is deliberately not gated either. API pricing lands at $10 per million input tokens / $50 per million output tokens, roughly 2.5× GPT-5.6 Sol on paper, though OpenAI argues smarter token usage makes it cheaper per task. That pricing happens to match Claude Fable 5 and 5.1 exactly — the first time the two labs' flagships have been at parity.

Two technical shifts are worth knowing because they'll define debates for months.

1. “Recurrent depth” — reasoning you can't watch. Astra uses a new reasoning technique whose internals obscure some or all of the model's chain-of-thought. That's great for speed and for users (no more lying-by-incorrect-“thinking”), but it is a genuine monitorability concern for safety researchers — if you can't see the reasoning, you can't audit it. Expect this to be the AlexNet-moment topic for the alignment community.

2. Codex notes across context windows. Also new: Codex can now preserve and retrieve context when the 1.05M-token window fills, replacing messy compaction with searchable notes across earlier windows. It's an experimental flag today, default tomorrow — and it quietly addresses the single most annoying failure in long agentic sessions: losing why a fix failed.

Four threads, in order of importance:

• The “multi-day autonomy” claim. Wharton's Ethan Mollick, who had early access, says he built a full simulation of the Library of Alexandria — a walkable, navigable, multi-day autonomous project. Unverified, but if it survives scrutiny, it's a materially new capability: sustained work across sessions, not one long context.
• The prime-gaps result. OpenAI paired the launch with a real math proof improving an 80-year-old bound — a signal it intends math to be a demo floor, not a demo ceiling.
• Competitive gravity. With Fable 5.1 leading the Intelligence Index and Astra leading specific agentic evals, “who's best” now splits by workload — speed-then-correctness (Fable) vs. hardest-end-to-end (Astra).
• Humanoid robots. OpenAI has separately confirmed it's building its own humanoid — meaning “computer use” quietly extends toward “physical use.”

Facts in this piece were compiled from OpenAI's official GPT-6 Astra announcement and developer documentation, Wikipedia, CNBC, Reuters, The Verge, Wired, Axios, DataCamp, the ARC Prize Foundation, Simon Willison's benchmark read-through, and ethics/tech journalism contemporaneous with the September 3–5, 2026 launch. Benchmarks are self-reported by OpenAI unless noted. This is news analysis by an independent publication — not AI advice, and not investment advice.