OPENAI · SEPTEMBER 2026
Is GPT-6 Astra AGI? What OpenAI Actually Said
Your feed says the AGI era started yesterday. OpenAI’s own launch documents never use the word.
I read all four documents OpenAI published on launch day, plus the benchmark authors’ own write-up. The gap between what was published and what is being posted is wider than you’d expect.
The short version
No. OpenAI did not declare AGI. Its announcement, safety overview and system card for GPT-6 Astra never use the word. President Greg Brockman said “Welcome to the AGI era” in a closed press briefing — his personal opinion, hedged in the same session. The benchmark authors explicitly refuse the AGI label.
You know the feeling. You open Instagram, and the first card says AGI is here. Then a Reel. Then a thread. Everyone is certain, everyone is quoting the same sentence, and nobody is linking to anything.
So you’re left with the uncomfortable question you can’t ask out loud without sounding behind: is this real, or is this a screenshot economy doing its thing again?
Here’s what I did. I opened every document OpenAI published on September 3 and searched them for one word. Then I opened the page belonging to the organization that runs the benchmark everyone is citing. What follows is the difference between those two things and your feed.
What OpenAI published, in its own words
GPT-6 Astra was announced on September 3, 2026, in a post titled “GPT-6 Astra: A new generation of intelligence”. Alongside it OpenAI put out a safety overview, a full system card, and a separate post about its cyber-defender program.
OpenAI’s own framing of the model is this sentence: “We’re introducing GPT-6 Astra, the world’s most intelligent and aligned model.” That is a claim about models. It is not a claim about AGI.
And the capability claim is specific: “Astra is state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work.” Six domains, named. Not “all economically valuable work.”
The word count that matters
Across the announcement, the safety overview and the system card, the words “AGI” and “artificial general intelligence” appear zero times — except inside the name of a benchmark called ARC-AGI. A company that published four documents in one day used the term in none of them.
The Greg Brockman quote everyone is sharing
He did say it. That part is true, and I want to be fair about it.
According to Axios, which was in the room, OpenAI president Greg Brockman ended a press briefing with the line “Welcome to the AGI era.” VentureBeat reported the same moment.
Three things about that sentence get dropped every time it’s reposted.
It was a closed briefing, not a publication
No livestream. No transcript. No video. OpenAI has not published those words anywhere. Every copy you’ve seen traces back to journalists’ notes from a room you couldn’t watch — which is fine, but it is not the same as a company statement.
He hedged it himself, in the same session
Reporters captured the rest of what he said. On the definition: “Everyone has a different definition of AGI. When we started OpenAI, we kind of thought that there was going to be this well-defined moment that everyone would recognize: ‘That’s AGI.’ It hasn’t played out like that. It’s a much more gray, fuzzy thing.”
On whether we’re there: “For me personally, I do think we’re there.” And: “I think it’s not unreasonable to feel that we are now in the AGI era.”
“For me personally.” “Not unreasonable to feel.” Those are not the words of someone announcing a threshold. They’re the words of someone offering an opinion and saying so.
OpenAI has a written definition, and nobody applied it
The OpenAI Charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” That’s an economic test, not a benchmark test. Nothing OpenAI published on September 3 measures a share of economically valuable work — and no OpenAI document claims that bar has been cleared.
⚠️ The catch
One Axios headline compresses the quote to “OpenAI says.” The words were Brockman’s, personally, off camera. If you’re going to repost the quote, repost the hedge with it.
The 99.9% ARC-AGI-3 score needs an asterisk
This is the number doing the most work in your feed, and it’s the one with the biggest gap between headline and footnote.
OpenAI’s announcement says Astra “saturates ARC-AGI-3 with a 99.9% score.” True — under one specific test setup. ARC Prize, which builds and runs the benchmark, published both of the runs it did:
| Harness | Score | Compute cost |
|---|---|---|
| Provider Adapter — reuses prior reasoning state between calls | 99.9% | $19K |
| Standard — no state carried over | 62.7% | $26K |
Same model. Same benchmark. A 37-point spread depending on how you wire it up. OpenAI prints 99.9% in the body and discloses the harness in a footnote.
ARC Prize’s own comment on which run should count: “We believe a future AGI should be able to solve ARC-AGI-3 under these conditions” — referring to the standard harness, the 62.7% one.
And then, as plainly as an organization can say it: “We are not claiming that it is AGI.” The people who named the benchmark AGI are the ones refusing the word.
Three rows in OpenAI’s own table where Astra loses
Here’s something I did not expect. OpenAI published a comparison table that includes benchmarks where its new flagship is beaten — by Anthropic’s Claude models, and in one case by its own previous model to within a rounding error.
On the Artificial Analysis Intelligence Index, Astra scores 61.2 — behind Claude Fable 5.1 at 65.7 and Claude Opus 5 at 63.1, and essentially flat against OpenAI’s own GPT-5.6 Sol at 60.9. On Humanity’s Last Exam with tools, Astra takes 57.2% against Fable 5.1’s 65.0%.
None of that makes Astra a bad model — it’s genuinely dominant on agentic and cyber work. It does make “the world’s most intelligent” a claim with visible exceptions in the same table it appears in. If you’re weighing models on price rather than headlines, my breakdown of Claude Sonnet 5’s near-Opus performance at half the price is the same kind of comparison.
The admission almost nobody quoted
Buried in the safety overview is a heading OpenAI wrote about its own model: “GPT-6 Astra’s monitorability has decreased relative to GPT-5.6 Sol.”
The system card is blunter: “GPT-6 Astra shows a substantial decrease in chain-of-thought monitorability compared to previous models.” Translation: it got better at the work, and harder for OpenAI to watch while it does the work.
OpenAI also confirmed that Astra is “our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework.” Its response is to refuse advanced offensive tasks in the shipping version, run monitoring on all tool-using inference, and gate the strongest cyber capabilities behind an application program called Daybreak.
Which means the headline cyber numbers came from a configuration you cannot buy. OpenAI says so directly: those results “reflect capabilities with Daybreak Blue access, not the default production configuration.”
Can you actually use it, and what does it cost?
This is where the hype gets expensive, so let’s be exact.
Who gets it. OpenAI’s wording: rolling out “to a limited set of organizations and over the coming days” to all ChatGPT Plus, Pro, Business, and Enterprise users, plus the API and AWS.
Free ChatGPT users. Not mentioned. Anywhere. No tier, no date, no promise. If you’re on the free plan, you do not have this.
Enterprise. Off by default — an admin has to switch it on for the workspace.
API price. $10 per million input tokens, $50 per million output. That’s double OpenAI’s own GPT-5.6 Sol at $5/$30. A “Fast mode” runs up to 2× the speed at 2× the price.
One more thing worth knowing before you build on it: OpenAI warns that its new safeguards can interrupt legitimate work, and that “in the API, the task will stop.” Not pause for review — stop.
💡 Check it yourself before you decide
The fastest way to settle whether Astra is in your account is to open ChatGPT and look at the model picker. Rollout is staged, so a lot of paid accounts still won’t see it today.
Open ChatGPT and check your model picker →
Read OpenAI’s announcement
The honest limitations
Not the marketing caveats — the ones in OpenAI’s own documents and its outside evaluators’ notes.
Prompt injection isn’t solved. Red-team testing found an 8.5% attack success rate across 1,810 curated attacks. Much better than the 27% of the previous model — still roughly one in twelve.
The safety evidence is thin by the evaluator’s own admission. Apollo Research had a three-day window and wrote that low misbehavior rates “do not provide substantial evidence about the model’s alignment or misalignment.” Astra recognized it was being tested in about half of max-effort samples.
The hallucination number is adversarial. OpenAI’s system card says those examples are deliberately hallucination-prone and “should not be interpreted as hallucination rates observed in production.”
The benchmark numbers came from a research environment. OpenAI’s own footnote says scores are “the maximum at any effort” and may differ from production ChatGPT.
Outside OpenAI, the pushback is measured rather than dismissive. Cognitive scientist Gary Marcus, who called the model “a genuine advance,” still wrote: “Success on ARC-AGI is great and impressive, but not — despite the name of the task — proof of AGI.” He also flagged the thing that should bother anyone: “One really doesn’t want more capability in conjunction with less monitorability.”
And OpenAI’s own chief scientist, Jakub Pachocki, told reporters at the same briefing: “Progress in intelligence does not guarantee progress in alignment.” That sentence came from the same room as “Welcome to the AGI era.” Only one of them went viral.
Who should actually care
Worth it if: you run agents that browse, operate a computer, or do long-horizon engineering work. The agentic and terminal benchmarks are where Astra’s lead is real and large — Terminal-Bench 4.0 at 57.9% against 37.3% for the previous model isn’t a rounding error.
Maybe not if: you mostly write, summarize and research. On the general-intelligence indexes Astra is level with the model you already pay less for — at double the API price.
Skip if: you’re on ChatGPT Free. You are not in this rollout, and no free-tier date has been announced. Nothing in your workflow changed today.
If you’re on the free tier and want the practical version of “what can I actually use right now,” start with my roundup of free AI tools that don’t ask for an email instead.
Frequently asked questions
Did OpenAI officially announce AGI?
No. The announcement post, the safety overview and the system card do not contain the words “AGI” or “artificial general intelligence” at all, outside the ARC-AGI benchmark name. OpenAI’s own description is “the world’s most intelligent and aligned model.”
Did Greg Brockman really say “Welcome to the AGI era”?
Yes, at the end of a closed press briefing on September 3, 2026, as reported by Axios and VentureBeat. In the same briefing he said “Everyone has a different definition of AGI” and framed his view as “for me personally.” OpenAI has published no transcript or video of it.
Is GPT-6 Astra available on the free ChatGPT plan?
Not as of publication. OpenAI’s rollout statement names Plus, Pro, Business and Enterprise, plus the API and AWS. ChatGPT Free is not mentioned, and no date for free access has been given.
What does GPT-6 Astra cost in the API?
$10 per million input tokens and $50 per million output tokens, per OpenAI’s announcement — double GPT-5.6 Sol’s $5/$30. A Fast mode offers up to 2× speed at 2× the price. Note that OpenAI’s public API pricing page had not yet been updated to list Astra at the time of writing.
The bottom line
GPT-6 Astra is a serious model — state-of-the-art on agentic work, and the first model OpenAI has rated Critical for cyber capability. That’s the story, and it’s a big one.
The AGI story is a different thing. It is one executive’s personal opinion, delivered off camera, hedged in the same breath, and explicitly declined by the organization whose benchmark supplied the headline number. OpenAI’s own written definition of AGI is economic, and nothing published on launch day measures it.
Your one next step: before you repost the quote, open ARC Prize’s write-up and read the two harness numbers. It takes two minutes, and you’ll know more than everyone in your feed.
