OpenAI Didn't Say It Achieved AGI. Here's What GPT-6 Astra Actually Does.

OpenAI released GPT-6 Astra on 3 September. Within a day the headlines said AGI had arrived, and a week on most people I speak to have absorbed the headline without reading anything underneath it.

So I read the primary sources — OpenAI's announcement and the benchmark organisation's own write-up — and the gap between what they say and what was reported is worth twenty minutes of your time. Not because the model is overhyped. Because the genuinely impressive thing about it got buried under an argument about a three-letter acronym.

Key Takeaway: The word "AGI" does not appear in OpenAI's announcement. The claim came from its president speaking personally at a press briefing, and from Nvidia's CEO. On the headline benchmark, the same model at the same settings scores 98.6% with OpenAI's own scaffolding and 62.7% on a neutral one — both published, both real, measuring different things. For your business the practical changes are pricing, a cybersecurity capability that is deliberately restricted, and the fact that the newest model is not the best one at writing.

Written by Derek Chua, founder of Magnified Technologies. We build automation and data systems for Singapore SMEs, which means we read model releases for what changes on Monday, not for what changes in ten years.

Who Actually Said "AGI"

I read the full text of OpenAI's announcement. Every occurrence of "AGI" on that page is a benchmark name — ARC-AGI-1, ARC-AGI-2, ARC-AGI-3. As a claim about the model, it is not there.

What OpenAI actually wrote is that Astra is "the world's most intelligent and aligned model." That is a bold claim and worth arguing about. It is not the same claim.

The AGI framing has two identifiable sources. The first is Greg Brockman, OpenAI's president, at a closed press briefing — reported by VentureBeat as closing the session with "Welcome to the AGI era" and, when actually asked whether Astra qualifies, answering more carefully: "for me personally, I do think we're there." Note the hedge, and note which of those two is the considered claim. He also said, of the term itself, that "everyone has a different definition of AGI... it's a much more gray, fuzzy thing." Sam Altman, who runs the company, is not quoted saying anything equivalent in either OpenAI's announcement or the briefing write-up.

Three days later, Nvidia CEO Jensen Huang posted on X that "AGI has arrived." Nvidia sells the hardware every one of these models is trained on — by Huang's own count, Astra's training run used upwards of 100,000 Grace Blackwell GPUs. That does not make him wrong, but it is context a reader deserves.

A hedged personal remark plus an enthusiastic supplier became "OpenAI claims AGI" in the aggregate. That is how this works now, and recognising the pattern is more durable than any fact about this particular model.

One Benchmark, Two Scores, Thirty-Six Points Apart

This is the part that actually teaches you something, and it will repeat with every major release from here.

The headline number was ARC-AGI-3: 99.9%. ARC Prize, the non-profit that builds the benchmark, published their own analysis the same day, and they ran Astra twice.

On their Standard harness — a minimal, provider-neutral interface, the same for every model — Astra scored 62.7%, costing $26,000 to run. On their Provider Adapter harness, which lets the model use OpenAI's own context-management features to preserve its reasoning state between requests, the same model at the same reasoning setting scored 98.6%, for $17,000.

Same model, same settings, same week. Thirty-six points of difference, produced by the scaffolding alone. (The 99.9% that made the headlines comes from a third run at a different reasoning setting again — which is its own small lesson in how these numbers travel.)

Neither number is a lie, and nobody falsified anything. ARC Prize published both scores side by side with a full table. OpenAI's footnote 1 says Astra was run with "our responses API harness, which changes two settings" — retained reasoning and compaction — and links to the post explaining them. It adds that "the changes do not specifically target ARC-AGI-3," which is fair. What it does not mention is that the neutral harness produced a different number. What happened is subtler and more ordinary than a cover-up: the flattering number travelled, and the footnote stayed home.

ARC Prize are explicit about which number answers the AGI question: "We believe a future AGI should be able to solve ARC-AGI-3 under these conditions," they write of the Standard harness — the 62.7% one. And then, plainly: "we are not claiming that it is AGI."

What makes this worth your attention is not that someone got caught. It is that "the model scored X" is now an incomplete sentence. The harness, the scaffolding, the tooling around the model can be worth more than a third of the score. When a vendor quotes you a benchmark — and they will — the useful question is no longer what did it score but under what conditions, and what did the neutral run say.

The Finding Everyone Skipped

Here is what I think is genuinely remarkable, and I have not seen it covered anywhere.

ARC Prize tested roughly 500 members of the public to establish a human baseline — not puzzle experts, ordinary people — and measured how many actions each took to solve an unfamiliar environment. Astra used fewer actions than the median human on 96% of levels, and 51.7% fewer on average.

Read that again, because the benchmark name obscures it. Dropped into a system it had never seen, with no instructions, Astra worked out how the thing operates — faster than most of the roughly 500 people ARC Prize tested.

ARC Prize had predicted this would be the durable dividing line. Their words: they "hypothesized that action efficiency would remain a dividing line between humans and AI." It did not. They call the result "a major milestone worth celebrating" and "a noticeable step-function change in frontier model capabilities."

They are, in the same post, both the people saying this is a genuine leap and the people saying it is not AGI. Those positions are entirely compatible, and holding both is what literacy in this field now looks like. The reason they still say no: ARC-AGI-3's environments "have deterministic, closed-ended mechanics and goals. It does not represent the complexity and open-endedness of the real world." Your business is not a closed-ended environment with deterministic mechanics. That is precisely the gap.

What Changes for Your Business This Month

Four things, in order of how likely they are to affect you.

1. Newest does not mean best at everything, and the gap shows in writing. This is the finding most relevant to any business producing content, and it is the one nobody led with. On OpenAI's own published comparison table, Astra scores 61.2 on the Artificial Analysis Intelligence Index against Claude Fable 5.1's 65.7 — OpenAI published a third-party index on which its new model comes second. The pattern across the rest of that table is consistent: Astra is dramatically stronger where answers are verifiable — code, maths, computer use, exploits — and flat or behind where the standard is taste. Before you move your marketing copy onto it because of the headlines, run the same brief through both and read the outputs side by side.

2. Pricing, if you use the API. US$10 per million input tokens and US$50 per million output. Fast mode doubles both for double the speed. If your automation runs at volume, that is a real line item — and worth comparing against what you are paying now rather than assuming the newest model is the right default.

3. Access is off by default for business accounts. OpenAI's announcement states plainly: "Enterprise administrators can enable Astra for their workspace; access is off by default at launch." If you are on a Business or Enterprise plan and wondering why you cannot see it, that is why. It is an admin setting, not a rollout delay.

4. The cybersecurity capability is real and deliberately restricted. Astra is the first model OpenAI has designated "Critical" for cybersecurity under its own Preparedness Framework. It scored 100% on ExploitBench and discovered two previously unknown vulnerabilities during evaluation. The public version refuses to build proof-of-concept exploits; it will do secure code review and patching. For most SMEs this is not a buying decision, but it is a reason to take your own patching cadence more seriously — the same capability that helps defenders is now broadly available.

There is a fifth thing that is not a product change but is the most interesting line in OpenAI's own post. They report that Astra's written reasoning is harder to monitor than its predecessor's; those are their words, in the launch announcement, about their own model. A company disclosing that its new system is less transparent than the one before it is unusual, and if you are handing more autonomy to AI tools this year, that trade-off is the one to actually think about.

Working With Magnified

We read these releases properly because our clients ask us whether to switch, and "the headlines said AGI" is not an answer. Most of the time the honest advice is that the model you are using is fine and the bottleneck is somewhere else entirely — usually the data, or a process nobody has written down.

If you are trying to work out whether any of this changes what you have built, we offer a free 30-minute consultation. Bring the actual workflow; we will tell you if there is a reason to touch it.

Frequently Asked Questions

Is GPT-6 Astra actually AGI?

The organisation that builds the benchmark it topped says no — "we are not claiming that it is AGI." OpenAI's own announcement does not use the word. One OpenAI executive said he personally thinks so, hedged, at a press briefing. On ARC Prize's neutral Standard harness it scores 62.7%. The 99.9% came from OpenAI's own scaffolding at a different reasoning setting, so the two are not the same run measured two ways.

Should I switch my business from whatever AI tool I use now?

Not on the basis of these headlines. Astra is meaningfully better at code, computer use and mathematics, and on OpenAI's own published table it sits behind Claude Fable 5.1 on the Artificial Analysis Intelligence Index. Test it on your own actual work before moving anything.

Why do the same model's benchmark scores differ so much?

Because a score measures the model plus the scaffolding around it. Run at the same reasoning effort, ARC Prize's neutral harness and OpenAI's context-management harness produced 62.7% and 98.6% for the identical model. Always ask which conditions — and which settings — a quoted number came from.

Is it available in Singapore?

Through ChatGPT Plus, Pro, Business and Enterprise plans, and through the OpenAI API, Microsoft Azure and AWS Bedrock. If you are on a business plan and cannot see it, an administrator has to enable it — access is off by default. Data residency and regional routing vary by platform and matter if you are in a regulated industry; check before moving production workloads.

What is Project Astra then? I have seen that name elsewhere.

A different thing entirely. Google DeepMind has an unrelated research prototype called Project Astra, built on Gemini. Two companies, two products, one name. GPT-6 Astra is OpenAI's.

If you found this useful, you may also want to read our take on Claude Fable 5 and what it means for small teams, or our guide to AI agents clicking your ads.

Sources

• GPT-6 Astra: A new generation of intelligence — OpenAI, 3 September 2026
• OpenAI's GPT-6 Astra on ARC-AGI-3 — Greg Kamradt, ARC Prize Foundation, 3 September 2026
• "Welcome to the AGI era": OpenAI launches GPT-6 Astra — VentureBeat, 3 September 2026 (Brockman briefing quotes)
• Jensen Huang on X: "AGI has arrived" — 6 September 2026