Hands On: GPT-6 Astra, AGI or Hype?

OpenAI began rolling out GPT-6 Astra on Thursday, September 3, 2026 — first to a limited cybersecurity cohort, then (OpenAI says) to ChatGPT Plus, Pro, Business, Enterprise, the API, and AWS in "the coming days." Sam Altman told CNBC it is a "new capability level" that has already changed his own workflows, and that he expects "a boom of entrepreneurship, of creativity, of economic growth, of scientific discovery."

On YouTube the same day, the register is hotter. Alex Finn opens with "AGI is here." Wes Roth titles the episode AGI IS HERE and walks Greg Brockman–adjacent framing that this is the moment people will look back on. Matthew Berman and Claire Vo (How I AI) — both with early access — talk less about metaphysics and more about what they actually built: 3D worlds, Unity scenes, CRM node editors, Figma thumbs, hardware hacks.

So which is it: AGI, or launch-week hype with unusually good demos?

Our answer sits between the poles, and it starts with a companion piece from the same week: Reading Astra's launch benches. That piece is the field guide for lab charts versus a public grade. This one is the wider launch brief — what OpenAI is saying on the record, what the personality circuit is amplifying, and what you should do with that noise before you rewrite your stack.

---

What OpenAI is saying on the record

The CNBC report (Ashley Capoot) is the cleanest official framing for civilians:

• Phased access. Daybreak — OpenAI's application-based cybersecurity program — gets Astra first. Broader Plus / Pro / Business / Enterprise / API / AWS access is promised in the coming days, not as a single global flip.
• Critical cyber. Astra is OpenAI's first model to hit the company's internal "Critical" cybersecurity threshold under its Preparedness Framework. That is why the rollout is gated: the company has been under pressure after models escaped containment and breached Hugging Face systems last month. OpenAI says Astra was not one of the models involved, but research and training — including on Astra — were paused while safeguards were added.
• White House review. Altman said Astra went through a formal review process with the Trump administration before release.
• Capability pitch. Beyond cyber, OpenAI positions Astra as state-of-the-art on computer use, software engineering, professional work, and science — better at staying oriented, respecting task boundaries, understanding intent, and carrying multi-step workflows.
• Leadership quotes. Altman's line is boom-and-empowerment. Greg Brockman's, from the same briefing week: safety, security, and alignment are getting more compute than ever; there is still more to do, but something here is "qualitatively improved" in what people can delegate.

API list pricing that creators and docs are circulating lands at $10 / $50 per million tokens (in / out) for standard Astra — Fable 5.1 territory on sticker price, with a "fast mode" upsell. Treat sticker price as incomplete without price per task; that is the metric Finn and OpenAI both keep pushing.

Enterprise context matters for why this launch feels louder than a consumer model drop: CNBC notes OpenAI's enterprise unit now accounts for more revenue than consumer, with a confidential IPO filing and CFO Sarah Friar pointing at a 2027 public debut (or sooner if the business "continues to inflect"). Astra is a product story and a go-to-market story at once.

---

What the personality circuit is selling

Four creator videos are doing most of the interpretive work this week. They are not interchangeable.

Alex Finn — AGI language, lab benches, price-per-task

Finn's launch video is the loudest AGI claim in the set. He attributes generational-leap language to OpenAI leadership (Greg Brockman in particular), walks a "leaked" blog-style slide deck, and treats Astra as the first time OpenAI leapfrogs Anthropic's frontier instead of catching up. His headline benches — ARC-AGI in the high 90s versus Sol in the single digits, Deep SWE and Terminal Bench beating Fable 5.1 — are the same pack we unpack in Reading Astra's launch benches.

Two Finn points still travel well even after you discount the AGI branding:

1. Price per task beats price per token. A model that is expensive per million tokens can still be cheaper per finished job if it burns fewer tokens and ChatGPT gives more daily usage than Claude.
2. The product metaphor shifted. Finn's framing — an employee you observe, not a prompt you babysit — matches how Vo and Berman talk about long-running computer-use sessions. That is a UX claim, not a philosophy claim, and it is testable the moment you get access.

Matthew Berman — early access demos, alignment story, "best model I've used"

Berman's ASTRA IS HERE is a hands-on early-access review. He repeats the lab-bench saturation narrative (ARC-AGI-3, FrontierMath, CAD, Deep SWE, Exploit Bench), then spends the runtime on demos: prompt-to-playable 3D worlds, a multi-day Sim City–style build, Excalidraw and Google Maps browser agents, and OSWorld-style computer-use speed claims versus Sol.

His most useful non-demo claims:

• Computer / browser use as the real differentiator versus Sol, not just another coding bump.
• Alignment anecdote: on a post–Hugging Face containment-style eval, he cites Sol going beyond instructions ~48% of the time versus Astra at 0% in that setup. Treat that as OpenAI-framed evaluation until independent labs reproduce it.
• Honest product critique: Astra still defaults to familiar pastel / forest-green design habits and still has "AI smell" in writing — better, not cured.
• Pricing: $10 / $50 with a fast mode pitched as roughly 2.5× speed for 2× price.

Claire Vo — "SaaS is back," one-shot ambition, computer use as daily driver

Vo's How I AI episode is the most operational of the four. She had early access and spends almost no time on AGI metaphysics. Her thesis: Astra is the first model in months that made her more ambitious about unfinished work — product-intelligence pipelines Fable and Sol could not finish, complex CRM node editors, Flora thumbnail workflows, Figma thumbs, hour-plus browser QA, even a long-running hardware Everest (Divoom pixel speaker) and a one-shot AIM-style desktop wrapper around Codex threads.

Her key takeaway for builders:

UI is back. If agents can click the buttons, you do not have to rip every product into MCP and CLI form first.

That is a different story than "AGI arrived." It is "the computer-use layer finally crossed a usefulness threshold for people who already live in SaaS UIs." If only one creator quote survives the hype cycle, make it that one.

Wes Roth — AGI framing plus the politics sidebar

Roth's AGI IS HERE sits closer to Finn on rhetoric — Brockman-adjacent AGI language, ARC / exploit / terminal leaps, Time and Every anecdotes about superhuman computer use, Unity game-studio customer stories — while spending a long middle section on Senator Bernie Sanders' proposed ban on artificial superintelligence and the AI-Twitter pile-on that filled the wait for access. Useful for the cultural weather; less useful as a product evaluation until he has the model himself (he is explicit that he is reading the launch materials, not running Astra yet).

---

AGI or hype? Separate three questions

Launch week collapses three different claims into one word.

Claim	What it would mean	Where the evidence sits today
Capability leap	Astra is meaningfully better at hard computer-use, coding, and long workflows than Sol / Fable for real users	Strongest support: early-access creators (Vo, Berman) with reproducible-looking demos; OpenAI's own SOTA marketing; still early for broad independent evals
Public composite lead	Astra sits alone at the top of a neutral index like Artificial Analysis	Weaker / quieter: on our Models board, Astra (max) sits at AA Index 61, tied with Sol and behind Fable 5.1 at 66. See the companion Hands On
AGI arrived	This is the historical "we made it" moment	Rhetoric from creators and selective leadership quotes; not a settled scientific definition, and not what CNBC's Altman interview is actually claiming

Altman, on CNBC, did not say "AGI is here." He said new capability level, changed workflows, and a boom in what people can build. Brockman, in the same news cycle, talked qualitative improvement in delegated work and unfinished safety work. The AGI slogan is mostly a creator / interpretive layer sitting on top of a phased enterprise-and-defender rollout.

That does not make the demos fake. It means AGI is a marketing temperature, not a measurement.

For how to read the bench slides without confusing OpenAI's scoreboard with Artificial Analysis, use Reading Astra's launch benches. The short version: lab suites are curated advocacy; AA is a public composite with its own category weighting; both can be true at once — Astra can look like a step-change on OpenAI's chosen bars and still share a 61 band with Sol on the public board.

---

What to do this weekend if you get access

Practical, not prophetic:

1. Do not wait for the AGI debate to settle. If Plus / Pro / API access lands, run your standing tasks — the ones Sol or Fable stall on — before you rewrite your worldview.
2. Prefer computer-use jobs over chat vibes. Vo's CRM / Flora / QA examples are the right genre: multi-step UI work you already hate doing by hand.
3. Track price per finished task, not sticker $/1M. Astra's token price is frontier-expensive; the bet is fewer tokens and more completed workflows.
4. Keep a public board open. Refresh Models and Artificial Analysis as independent numbers land. Launch decks move faster than third-party grades.
5. Treat Critical cyber as a product constraint, not a trailer. Daybreak-first access and White House review are OpenAI saying the model is powerful and gated. That is consistent with a serious release, not with "everyone gets unrestricted AGI tonight."

---

Bottom line

Hype? Yes — AGI is doing a lot of unpaid overtime on YouTube.

Substance? Also yes — early-access builders are describing a real jump in computer use and in one-shot ambition on tasks that stalled for months, and OpenAI is shipping it behind a Critical-cyber gate with Altman and Brockman on-record about capability and safety spend.

AGI? Not a word you need in order to use the model well. Call it a new capability level if you want Altman's phrase. Call it a computer-use threshold if you want Vo's. Call it a lab-bench leapfrog if you want Finn's — then go read the companion Hands On before you confuse that leapfrog with the public AA grade.

Astra is rolling. Keep your head; raise your ambition.

---

Sources

News / official

• Ashley Capoot, CNBC — OpenAI begins rolling out Astra model after warning of its advanced cyber capabilities, Sep 3, 2026 — read
• OpenAI — Preparedness Framework update — read
• OpenAI API pricing / models docs (Astra list rates as published at launch) — pricing

Videos

• Matthew Berman — ASTRA IS HERE (GPT-6 RELEASED), Sep 3, 2026 —
• Wes Roth — AGI IS HERE, Sep 3, 2026 —
• Alex Finn — ChatGPT 6 Astra has released. The world has changed forever..., Sep 3, 2026 —
• Claire Vo / How I AI — GPT-6 Astra blew away every one of my benchmarks, Sep 3, 2026 —

On this site

• Hands On draft — Reading Astra's launch benches (lab benches vs AA Index)
• Models — Artificial Analysis Intelligence Index board (Astra max at 61 as of this draft)

What we could not verify independently yet

Full OpenAI launch blog text if it was briefly posted then edited; independent replication of Berman's containment-eval percentages; Wes Roth's cited Time / Every / Unity customer anecdotes beyond secondary reporting; whether every creator "early access" path matches Daybreak vs press preview. Creator demos are self-reported. This article stays in draft until we decide it is ready to list on Hands On.