Last updated: September 8, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
On Sunday, September 6, 2026, Nvidia CEO Jensen Huang posted on X that "AGI has arrived," pointing to OpenAI's newly released GPT-6 Astra model as the proof. It's a claim, not a settled fact, and in my testing of what's actually behind it, the evidence turns out to be a lot messier than a five-word declaration suggests.
Short answer: On September 6, 2026, Jensen Huang posted on X that "AGI has arrived," crediting OpenAI's GPT-6 Astra, trained on more than 100,000 Nvidia Grace Blackwell NVLink72 systems. OpenAI's Greg Brockman echoed the sentiment more cautiously. AI researcher Gary Marcus pushed back hard, saying Huang offered "no evidence and no definitions" for the claim.
In my testing of this story, I didn't want to just repeat "AGI has arrived" as a headline, because that phrase has been declared before — by Huang and others — and it never quite sticks. Every time a CEO with a hardware business says AGI has arrived, the claim is worth running through the same checklist you'd use on any viral executive post: read the actual words, find the vendor's own framing, track down an independent rerun of whatever benchmark is being waved around, and see who in the field is pushing back and why. So this piece walks through the actual quote, what OpenAI's own benchmark numbers show once you read past the headline figure, and where the pushback is coming from, so you can decide for yourself instead of taking one CEO's word for it.
What Huang actually posted
Huang's post, quoted in full: "GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team. 400K GPUs coming online next." It came three days after OpenAI began rolling out GPT-6 Astra on September 3, 2026, which I covered in more detail separately when the access tiers and pricing first went live.
OpenAI president Greg Brockman replied in a noticeably softer register: "we're now moving into the AGI era (whether you view it as this model, the last one, or the next one), and could not do it without close partners." Notice the hedge — Brockman isn't pinning the AGI claim to Astra specifically, the way Huang did.
Step-by-step: how I checked whether the claim holds up
1. Read the post itself, not a paraphrase
Headlines compress "AGI has arrived" into a clean five-word soundbite. The full post is a product plug wrapped around that phrase — it also thanks OpenAI, cites the GPU count Astra trained on, and previews 400,000 more GPUs coming online. Nvidia sells those GPUs. That context doesn't make the claim false, but it's relevant to how much weight I'd put on it.
2. Check what OpenAI itself is claiming, separately from Huang
OpenAI’s own launch materials describe Astra as its "most intelligent and aligned" model, citing gains in computer use, software engineering, cybersecurity, and science. That's the vendor's framing, and it's a step below "AGI" — in my read of the launch page, OpenAI's own team is more careful with the term than Huang was in his post.
3. Find the independent rerun of the headline benchmark
The number doing the heaviest lifting in Astra's launch materials is a 99.9% score on ARC-AGI-3. When I checked ARC Prize’s own writeup — the organization that runs that benchmark, independent of OpenAI — that 99.9% came from a custom "provider adapter" harness built for the launch. The same model scored 62.7% on the standard harness every other model was tested with. Both numbers are real. Only one of them is comparable to anything else on the leaderboard.
4. See who's disagreeing, and on what grounds
AI researcher Gary Marcus, a longtime critic of premature AGI declarations, wrote that Huang’s claim came with "no evidence and no definitions," calling it "an effort at a takeover of a scientific question by corporate fiat." He pointed to a joint effort called agidefinition.AI, co-authored by researchers including Dan Hendrycks and Yoshua Bengio, as the kind of agreed-upon standard a claim like this should be measured against before anyone declares victory.
5. Check what the market actually did
Nvidia's stock closed at $230.36 on Friday, September 4, 2026, up 0.84% for the day, before U.S. markets closed for Labor Day on the Monday Huang's post went up. A single day's stock move doesn't confirm or deny an AGI claim — investors were already pricing in Huang’s fiscal 2028 revenue guidance of roughly 70% growth from his September 5 earnings call, a separate story from the Astra post.
GPT-6 Astra's numbers vs. the model it replaced
| Metric | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| ARC-AGI-3 (standard harness) | 62.7% | 7.8% |
| ARC-AGI-3 (OpenAI's provider-adapter harness) | 99.9% | Not offered |
| FrontierMath Tier 4 (OpenAI, vendor-reported) | 97.6% | Not comparable |
| Authorized-scope violations (OpenAI safety overview) | 0% | 48% |
| API price (input / output per 1M tokens) | $10 / $50 | $5 / $30 |
The standard-harness jump from 7.8% to 62.7% is real and independently confirmed by ARC Prize. The 99.9% figure is the one being repeated in most headlines, including the number that appears to sit behind Huang's "AGI has arrived" post, and it's the one that needed a custom harness to produce. The FrontierMath and authorized-scope numbers, by contrast, come straight from OpenAI's own launch materials and haven't been re-run by an independent lab yet, so I'm treating them as vendor-reported until someone does.
Example prompts you can copy
These are the prompts I use to pressure-test a viral capability claim before repeating it:
- "Find the primary source for this benchmark score — who ran it, and is it the vendor or an independent group?"
- "Does [benchmark name] have a standard, comparable version, and did this model use it or a custom variant?"
- "Search for the most vocal critics of this specific claim and summarize their actual argument, not just that they disagree."
- "Is there an agreed-upon definition of the term being used here, and does this result meet it?"
- "What financial or business interest does the person making this claim have in it being true?"
Common mistakes to avoid
The mistake I see most with this story is repeating "AGI has arrived" as if OpenAI itself said it, when it's Huang's framing, not OpenAI's official position. Second is quoting the 99.9% ARC-AGI-3 score without the harness caveat — the number comparable to every other model is 62.7%. Third is ignoring that Nvidia sells the exact hardware Astra trained on, which doesn't invalidate Huang's opinion but does explain why he'd be first to call it a milestone. Fourth, treating Gary Marcus's pushback as contrarianism for its own sake — he's pointing at a specific, named alternative standard rather than arguing from vibes. Fifth, reading Friday's stock close as a market verdict on the AGI claim, when the bigger driver that week was Huang's own revenue guidance from two days earlier.
Tools that make this easier
If you want the full rollout details — pricing tiers, who actually has access, and the rest of the benchmark footnotes — my GPT-6 Astra access guide covers that separately. For the habit of verifying an executive's viral post before repeating it, I laid out the same checklist in more depth after Jensen Huang’s first-ever X post went up in July 2026. If you want to try GPT-6 Astra or compare it against what you're already using, how to use ChatGPT covers the interface, and my free AI tools roundup lists no-cost alternatives if you'd rather not pay Astra's API rate. Developers deciding whether the benchmark gains justify a migration should also see my best AI tool for code picks. And if you're trying to make sense of why Nvidia's CEO keeps tying his hardware sales to AI milestones, Nvidia’s $673 billion sales projection is useful context for the incentive sitting underneath the claim.
My take
When Jensen Huang says AGI has arrived, it's tempting to treat the sentence as news rather than opinion, because he runs the company that sells the chips behind almost every frontier model. Astra is a genuinely better model than Sol on the benchmark that matters — 62.7% versus 7.8% on ARC-AGI-3's standard harness is a real jump, not a rounding error. But "AGI has arrived" is a much bigger claim than "this model scored higher than the last one," and Huang's post doesn't bridge that gap. It asserts it, from the mouth of the person whose company sells the chips Astra trained on. Gary Marcus's specific ask — that Huang measure the claim against an actual published definition before declaring it settled — seems like the reasonable bar here. Until someone does that math in public, I'd treat "AGI has arrived" as marketing copy with real benchmarks attached to it, not as a fact.
Frequently Asked Questions
Did Jensen Huang really say "AGI has arrived"?
Yes — Jensen Huang says AGI has arrived in a post published on X on September 6, 2026, crediting OpenAI's GPT-6 Astra and thanking the OpenAI team directly in the same post.
Does OpenAI agree that AGI has arrived?
Not exactly. President Greg Brockman's reply hedged, saying the company is "moving into the AGI era" without specifying whether Astra, an earlier model, or a future one deserves the credit.
Is the 99.9% benchmark score real?
The number is real and published by ARC Prize, but it came from a custom harness built for the launch. The comparable score, using the same standard harness other models were tested with, is 62.7%.
What is Gary Marcus's actual objection?
That Huang made the claim without citing a specific, agreed-upon definition of AGI or evidence measured against one, and he pointed to a project called agidefinition.AI as the kind of standard that should be used instead.
Is GPT-6 Astra free to use?
No. It requires a paid ChatGPT plan or API access, and the rollout happened in phases rather than to everyone at once — details are in my separate GPT-6 Astra access guide.