OpenAI unveiled GPT-6 Astra last Thursday and described it on its website as 「the world’s most intelligent and most aligned model.」 The company did not say it had achieved AGI.
NVIDIA CEO Jensen Huang quickly raised the temperature. In a post on X, he wrote: 「AGI has arrived. Congratulations to the OpenAI team.」 That statement landed harder because OpenAI’s official launch page did not contain any equivalent declaration.
OpenAI left room for interpretation
OpenAI’s own language was much less definitive than Huang’s.
During a press call on launch day, OpenAI president Greg Brockman said, 「Welcome to the age of AGI.」 He went on to say that if people looked back a few years from now and asked when AGI was born, 「it’ll probably be around this time, maybe this model.」
Then he added two qualifiers. First, he said it was his personal judgment. Second, he said AGI is a gray and fuzzy concept and that users can decide for themselves whether this counts.
That left OpenAI without a formal, public standard for saying AGI had arrived. The ambiguity was already visible before Astra launched. Sam Altman had previously criticized the term AGI in public, saying it was badly defined and close to a marketing term. So the company’s position now looks split: Altman has questioned the usefulness of the label, Brockman used it while stressing it was a personal view, and the launch page avoided a direct AGI claim.
ARC-AGI-3 produced two very different scores
The dispute turned on one benchmark: ARC-AGI-3.
OpenAI said Astra scored 99.9% on ARC-AGI-3 and effectively saturated the benchmark. That number became a centerpiece of the release.
ARC Prize Foundation pushed back the same day and said the result should not be treated as a standard AGI result. Its explanation was straightforward: Astra had been tested in two different environments.
On the ARC-AGI-3 leaderboard, Astra’s results were listed separately by setup. In the same ranking, Claude Opus 5 scored 30.2% and GPT-5.6 Sol scored 7.8%.
In OpenAI’s deeply customized environment, which ARC Prize said allowed the model to preserve reasoning state and use proprietary techniques to manage long conversations, Astra did reach 99.9%. The cost of that run was nearly $19,000.
In the standard environment used equally across vendors, without those custom tools, Astra’s top score was 62.7%. ARC Prize said the cost there rose to $26,000.
The foundation’s position was clear: if AGI is to mean anything, a system must solve problems under standard conditions. In that framing, a 99.9% score with tools and a 62.7% score in a standard setup are not interchangeable.
ARC Prize still credited Astra with meaningful progress. It said the model showed a 「step-function」 improvement in generalization, and on 96% of levels Astra used fewer actions than humans, with an average reduction of 51.7%. According to the foundation, this was the first time AI had beaten humans across the board on action efficiency.
Even so, ARC Prize repeated a point it had made when ARC-AGI-3 launched: maxing out this benchmark does not prove AGI. Its reasoning was that even a difficult benchmark remains a closed environment with fixed mechanics and bounded objectives, which is not the same as the open complexity of the real world.
Huang’s post carried three key numbers
Huang’s full post said: 「GPT-6 Astra, trained on approximately 100K NVIDIA Grace Blackwell NVLink72s. From ChatGPT to o1 to Astra in just 4 years. AGI has arrived. Congratulations OpenAI team. 400K more GPUs coming online next.」
Three numbers stood out: 100,000, four years, and 400,000.
The first was 100,000. Huang said Astra was trained on about 100,000 NVIDIA Grace Blackwell NVLink72 systems, which the source article described as OpenAI’s largest training run so far.
The second was four years. In one compressed timeline, Huang linked ChatGPT, o1, and Astra, connecting conversational models, reasoning models, and what he called AGI.
The third was 400,000. He added that 400,000 more GPUs were coming online. The post did not say whose GPUs those were, what exact configurations they referred to, or what workloads they would support.
Placed together, those numbers created a powerful commercial narrative: if roughly 100,000 GPUs helped produce Astra and Astra is being framed as AGI, what might another 400,000 GPUs produce? In that reading, AGI becomes not only a technical milestone but also a justification for continued demand for massive compute.
Public claims are not contract recognition
The stakes are higher because AGI is not just a rhetorical label. It appears in commercial agreements.
According to the source article, a new OpenAI-Microsoft agreement signed last October said that even if OpenAI declares AGI, an independent expert panel must still verify it.
In February this year, the two companies jointly reaffirmed that the contract’s AGI definition and recognition process remained unchanged.
Then in April, the agreement was revised again so Microsoft’s revenue share became unrelated to OpenAI’s technical progress. The article said outside observers widely read that change as a sharp weakening of any commercial trigger tied to the AGI clause.
That means Brockman’s comments on a press call, Huang’s declaration on X, and Astra’s benchmark scores do not amount to AGI recognition in the contractual sense described in the article.
The same source said OpenAI is at a critical stage in IPO preparations, while Anthropic is also moving toward a listing. It also said the term AGI already appears in investment agreements involving OpenAI, Microsoft, and Amazon, giving the word weight beyond technical belief alone because it touches equity pricing and commercial control.
That helps explain why Brockman framed his remarks as a personal judgment and why OpenAI’s website avoided a formal AGI declaration. Public enthusiasm and legal wording are not the same thing.
No one has final authority to declare AGI
The source article also cited a Bloomberg feature that walked through the basic question of who gets to decide when AGI has arrived. Its answer: no one.
The definitions differ sharply.
OpenAI’s version is a system that outperforms humans on most economically valuable tasks.
A 2023 Google DeepMind paper focused less on economic value and more on generality, asking whether a system can learn new skills from very little data.
ARC Prize took a blunter line, arguing that economic value is the wrong yardstick for intelligence. Its definition is a system that can efficiently acquire new skills beyond its training data.
The labels are drifting too. Anthropic’s Dario Amodei prefers 「powerful AI.」 Microsoft’s Mustafa Suleyman has used 「artificial capable intelligence.」 Meta and OpenAI have lately leaned more toward 「superintelligence.」
The timelines are no cleaner. In September 2024, Altman said systems pointing toward AGI were beginning to emerge and said superintelligence could arrive within 「thousands of days.」 A year later, after GPT-5 launched, he said 「we are missing something quite important,」 without specifying what that was. Another year later, Brockman said at the Astra launch, 「Welcome to the age of AGI.」
At DeepMind, Shane Legg has since 1999 put the probability of AGI arriving before 2028 at 50%. Demis Hassabis has said it may still take five to 10 years, adding that the timeline appears shorter only because the definition of AGI is being diluted.
The picture that emerges is fragmented. Model companies control the product story and publish benchmark numbers without always making a formal AGI claim. Chip companies control the compute story and may be quicker to say the era has arrived. Benchmark organizations hold the measuring stick and refuse to move the goalposts. Capital partners such as Microsoft stick to the contract language.
For the past four years, the industry argument centered on who would build AGI first. More recently, the tense has shifted from 「one to three years away」 or 「maybe in six months」 to 「it is already here.」 What still does not exist is a broadly accepted standard for recognizing it.
Referenced sources in the original article
- Jensen Huang’s X post: https://x.com/JensenHuang/status/2096700264569090384
- Bloomberg feature: https://www.bloomberg.com/news/features/2026-09-04/what-is-agi-openai-anthropic-race-for-artificial-general-intelligence
- Axios report: https://www.axios.com/2026/09/03/openai-astra-gpt-6-agi-brockman
The original piece said it was sourced from the WeChat public account Xinzhiyuan, written by ASI Qishilu and edited by Yuanyu.
