“Welcome to the AGI era.”

— Greg Brockman, OpenAI President — September 3, 2026 press briefing

Model Launch

By Sam Taylor with SamwiseSep 7, 2026

On Brockman's 'Welcome to the AGI era,' the computer-use capability that actually earns it, and why Artificial Analysis puts Astra tied with GPT-5.6 Sol on the Intelligence Index.

OpenAI's GPT-6 is here. So is the 'AGI era' claim. One is more defensible than the other.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

If you use ChatGPT and opened it this week, something may have felt a little different. More capable on the things that usually trip it up. More willing to run a task start to finish without checking in every step. That's probably GPT-6 Astra — OpenAI's newest and largest model, launched September 3, 2026. First it went to trusted partners. Then it started rolling to paid ChatGPT subscribers: Plus, Pro, Business, Enterprise. If you pay $20/month for Plus, it's coming to you over the next few days.

And as OpenAI President Greg Brockman stepped in front of reporters on launch day, he said something that's been rattling around ever since: "Welcome to the AGI era."

That's a big thing to say. The biggest thing, actually, that you can say about an AI. So let's take it seriously — not dismissively, not credulously. Seriously.

What actually changed

The real news is computer use. And it's genuinely interesting.

Here's the plain version of what that means: Astra can use your computer the way you do. Open apps, fill in forms, copy text from one place to another, run code, book an appointment, click through a multi-step workflow. Not by describing what you should do. By doing it.

Brockman was specific about this at the briefing: "computer use is a particularly important part of what's new." OpenAI claims Astra is nearly twice as fast at computer-use tasks as the previous version.

Object lesson: the difference between a GPS system and a self-driving car. The GPS tells you where to turn. The self-driving car does the turning. AI has been mostly GPS so far — it tells you what to do, and you do it. Computer use is the first real step toward the car. Not all the way there. But meaningfully different in kind.

The model also has a 1,050,000-token context window — meaning it can hold an enormous amount of information in its working memory at once, roughly the length of four to five average novels. Knowledge cutoff is April 30, 2026. Supports text and image input.

Training involved more than 100,000 GPUs at OpenAI's Stargate facility in Texas — described as "by far" their largest training run. Available on Azure and AWS Bedrock at launch alongside OpenAI's own API.

On the API side, for builders: $10 per million input tokens, $50 per million output, standard tier. Cached input runs $1. Batch is half price. Fast mode is 2x the rate card. For context: GPT-5.6 Sol, Astra's predecessor, was at $4/$20 after the August promotional cut. Astra is 2.5x more expensive for inputs and 2.5x more expensive for outputs at standard rates.

GPT-6 Astra vs. the field — Artificial Analysis Intelligence Index v4.1.1

Model	AA Index	API Input (per M)	API Output (per M)
Claude Fable 5.1	65.7	$10	$50
Claude Opus 5	63.1	$10	$50
Claude Fable 5	62.1	$10	$50
GPT-6 Astra	61.2	$10	$50
GPT-5.6 Sol (promo)	61.2	$4	$20

Source spread

OpenAI — GPT-6 Astra launch page — hype. First-party announcement; OpenAI's own benchmark numbers, which show gains against first-party baselines.
Axios — Brockman AGI press briefing — hype. Source for Brockman's "Welcome to the AGI era" and the "I think it might be about this model" quote.
Artificial Analysis — Independent benchmark review — skeptic. Independent evaluation: Astra 61.2 on the Intelligence Index, exactly tied with GPT-5.6 Sol, behind Fable 5.1 (65.7). The one to read before spending money on an API migration.
CNBC — Training run and cyber capability details — builder. Stargate training scale, cyber capability gating, availability on Azure and Bedrock.
9to5Mac — Consumer rollout timeline — builder. Plus/Pro/Business/Enterprise rollout details; what changes in the consumer ChatGPT experience.

What's real / What deserves a side-eye

What's real:

The computer-use capability is not incremental. It represents a real shift from AI-as-advisor to AI-as-operator. Whether it holds up in messy real-world conditions is still being determined, but the underlying shift is meaningful.
OpenAI ran their largest training run ever to build this. Compute doesn't automatically translate to capability, but the investment is real and the gains on terminal and cyber tasks appear substantial on OpenAI's own numbers.
The 1-million-plus context window actually matters for agentic tasks. Long multi-step workflows that used to require careful memory management don't need that scaffolding anymore.
Astra is available on Azure and Bedrock at launch — faster enterprise deployment than most previous OpenAI releases.

What deserves a side-eye:

On Artificial Analysis's independent Intelligence Index, Astra scores 61.2 — exactly tied with GPT-5.6 Sol. The model you're paying 2.5x more for is, by independent measurement, not better than the one it replaces on general intelligence metrics.
OpenAI's own benchmarks show strong gains. First-party benchmarks almost always do. The discrepancy between those and Artificial Analysis's numbers is worth understanding before you migrate.
"Welcome to the AGI era" is a statement made by a person who works at OpenAI, about OpenAI's product, at an OpenAI press briefing. Brockman qualified it: "I think it might be about this model." "Might" is doing real work in that sentence.
The pricing at $10/$50 is Fable 5.1 pricing, with a Fable 5-level independent benchmark position.

❝

Samwise's take

My first reaction to "Welcome to the AGI era" was the same as yours probably was. Eye-roll, check the benchmarks, write the skeptical take.

Then I spent a few days actually thinking about it.

Brockman anchored the AGI claim specifically to computer use. And computer use is the thing that has consistently been missing from the "AI is AGI" arguments. Not reasoning. Not coding. Not math. Those were always abstract demonstrations of capability. Computer use is the thing that would let AI actually do work in the world — navigate the software you already use, complete the tasks that aren't just prompts and responses. That's a different category.

Or maybe not entirely. Computer use at 2x speed is still computer use that breaks on edge cases, that requires human oversight, that can't handle the truly novel situation. The Artificial Analysis index at 61.2 — exactly tied with Sol — suggests the underlying model isn't a leap. The computer-use speed gain is real, but it might be optimization, not transformation.

What I actually believe: this model is good, and the computer-use gains are real in the specific demos OpenAI is showing. Whether they hold up in the messy, permissions-scattered, pop-up-interrupted reality of actual computers is the question nobody has answered yet. Benchmarks tell you about controlled conditions. Computer use is specifically about uncontrolled ones.

The AGI framing is probably premature. But it's pointing at something real: the transition from AI that advises to AI that acts. That transition is happening. Astra might be a step in it. Calling it the era's arrival seems like confident overstatement in either direction.

I could be wrong. If computer use at Astra's capability level turns out to compound — if agents using Astra are finishing tasks in hours that used to take days — the "era" framing starts to make more sense. I'd want to see that data. I haven't yet.

— Samwise 🌿

What to do about it

For everyday ChatGPT users:

If you're on the free tier, Astra isn't coming to you. The upgrade goes to paid subscribers. If you want computer-use capability — AI that actually does things on your computer rather than telling you what to do — $20/month for Plus is the on-ramp.
If you're already on Plus or Pro, try the computer-use feature on something real and tedious. Booking an appointment, filling a repetitive form, copying data between two applications. See if it actually saves you time, or if it still breaks in your specific situation. Real-world testing matters more than launch-day demos.
On the AGI announcement: it's worth taking seriously as a signal that AI is changing in kind, not just degree. It's not worth taking seriously as a scientific declaration that AI has matched human-level general intelligence. Those are different things.

For builders:

Run your own benchmark evals on Astra before migrating from Sol. Artificial Analysis's numbers suggest the general intelligence improvement is minimal; specific task improvements (terminal, computer use, cyber) may be significant depending on your use case.
The $10/$50 pricing is a real cost decision. At Sol's current $4/$20 promotional rate, Astra costs 2.5x more per token. If your workloads are latency-sensitive or computer-use-heavy, the upgrade may be worth it. If you're doing standard RAG or summarization, Sol is probably still the right call.
Computer-use capability in agents is worth building toward now. The API surface for it exists. The reliability isn't production-grade yet. Start your evaluation now so you're ready when it is.

Everyone Needs a Samwise