OpenAI's GPT-6 Astra Sparks AGI Debate Following Huang's Announcement

Last Thursday, OpenAI released GPT-6 Astra.

The official website defines it as "the most intelligent and most aligned model in the world."

Immediately afterward, Jensen Huang posted on X, taking the matter to another level:

AGI has arrived. Congratulations to the OpenAI team.

In contrast, the official release page for GPT-6 Astra contains no mention of "AGI has been achieved."

A GPU seller made promises on behalf of those building models.

More interesting is the person who created the question.

The OpenAI page states that Astra achieved 99.9% on ARC-AGI-3, "saturating" this AGI-named benchmark.

The evaluation agency responsible for creating the test, ARC Prize, reviewed the benchmark scores but rejected them outright: "Not valid."

The ARC Prize Foundation posted a breakdown that day:

This score comes from OpenAI’s customized testing environment; when converted to a standardized environment that treats all vendors equally, it drops to 62.7%.

Their conclusion was only one sentence: Astra has made substantial progress in generalization, "but we do not claim it is AGI."

The chip supplier stamped directly, the model manufacturer left room for maneuver, and the benchmark institution responsible for setting the questions refused to sign.

This was probably the most surreal moment in the AI community over the past few days: AGI wasn’t truly “achieved” this time—it was merely “announced” by various spokespersons on social media.

What does OpenAI say itself?

During the press conference call on the day of the release, President Greg Brockman made a highly provocative statement:

Welcome to the AGI era.

He even predicted that if we look back in a few years to determine when AGI was born, “it will likely be this period, possibly this model.”

Does it sound like an official announcement?

But he immediately added two qualifications: first, this was his "personal judgment"; second, AGI is a gray and ambiguous concept, and "whether it qualifies is up to the user to decide."

At the launch of its most important product, the president used his "personal opinion" to define the era and even handed the final decision over to netizens.

This only proves one thing: Even within OpenAI, there is no clear, publicly verifiable AGI certification.

More interestingly, CEO Altman's attitude.

Before Astra's release, he publicly criticized "AGI" as a poorly defined term bordering on marketing jargon.

The CEO finds the term too vague; the president treats it as a "personal opinion," and the official website avoids mentioning it altogether.

Huang Renxun’s endorsement and congratulations on AGI landed precisely in this subtle gap.

How strong is Astra?

The OpenAI page states:

Astra achieved an astonishing score of 99.9% on the ARC-AGI-3 benchmark named AGI, saturating the test.

But the comeback was swift.

On that day, the ARC Prize Foundation, responsible for setting the questions, published a post publicly breaking down the "99.9%".

Astra originally took it twice.

On the ARC-AGI-3 leaderboard, Astra's results for two environments are listed separately; Claude Opus 5 scores 30.2%, and GPT-5.6 Sol scores 7.8%.

In OpenAI's deeply customized "tools-enabled" environment—which allows the model to maintain reasoning state and manage long conversations using proprietary technology—it achieved a score of 99.9% at a cost of nearly $19,000.

But if switched to a standardized "bare-metal" testing environment that treats all vendors equally, its highest score is only 62.7%, and the cost surges to $26,000.

The ARC Prize takes a firm stance: future AGI must be able to solve problems under standard conditions. Achieving 99.9% with tools versus 62.7% without them is entirely different.

The creator also acknowledges that Astra delivered a "leap-like" surprise, using fewer actions than humans in 96% of levels, averaging 51.7% fewer.

This is the first time AI has completely outperformed humans in action efficiency.

The ARC Prize had already made it clear beforehand: on the day ARC-AGI-3 launched, they explicitly stated that achieving the maximum score on this benchmark does not equate to "proving AGI has been reached."

The reason is simple: no matter how challenging a test environment is, it is still closed, with fixed mechanisms and limited objectives, and thus cannot represent the complexity and openness of the real world.

Just like a top student who scores full marks on an exam, they may not necessarily continue to excel in the real workplace.

If the questioner doesn't acknowledge it, why is Lao Huang the most eager to announce that "AGI has arrived"?

When you carefully break down Jensen Huang's post, you'll find its information structure is much more substantial than it appears on the surface.

The full post is:

GPT-6 Astra was trained on approximately 100,000 NVIDIA Grace Blackwell NVLink72 units. From ChatGPT to o1 to Astra, it took only four years. AGI has arrived. Congratulations to the OpenAI team. Another 400,000 GPUs are coming online next.

First number: 100,000.

Astra is the largest training run to date by OpenAI, utilizing over 100,000 GPUs.

Second number: 4 years.

From ChatGPT to o1 and then to Astra, Jensen Huang connected dialogue models, reasoning models, and so-called AGI along a single timeline.

The third number: 400,000.

Another 400,000 GPUs will be coming online.

Who owns these 400,000 units? What model are they? What are they used for? Huang Renxun left this as a mystery.

But when these three numbers are combined with "AGI has arrived" in the same post, readers will form this business narrative:

One hundred thousand GPUs were thrown into Astra (i.e., AGI)—what will the next four hundred thousand GPUs produce?

In this context, AGI is no longer an abstract technological endpoint, but rather the strongest proof of computational demand.

Once AGI is widely recognized as having been achieved, exorbitant computing power will no longer be a corporate expense, but a necessity that no one dares to walk away from.

In last October’s new agreement between OpenAI and Microsoft, it was stipulated that after OpenAI announces AGI, it must be verified by an independent panel of experts.

In February this year, both parties jointly reaffirmed that the definition and recognition process of AGI in the contract "remain unchanged."

By April, the agreement was revised again, with Microsoft’s revenue share changed to be “unrelated to OpenAI’s technological progress,” which the public widely interpreted as a significant weakening of the AGI clause’s impact on commercial interests.

In other words, what Brockman says at the press conference, what Huang Renxun posts on X, or the score achieved by ARC-AGI-3 do not constitute contractual recognition of AGI.

Right now is a critical moment for OpenAI as it prepares for its IPO, while its rival Anthropic is also racing toward going public.

The term "AGI" has long been included in OpenAI's investment agreements with Microsoft and Amazon, extending beyond technological belief to encompass equity valuation and business control.

This explains why Brockman only dared to call it "personal judgment" and why the OpenAI website strictly avoids mentioning AGI.

The public discourse can enthusiastically welcome the AGI era, but in contracts, not a single word more can be written.

A long article in Bloomberg walked through this issue, and the answer is that no one is in charge.

First, let's look at the definition.

OpenAI's version is "a system that outperforms humans on most economically valuable tasks."

Google DeepMind's 2023 paper focuses on generalizability and the ability to learn new skills with minimal data, rather than economic value.

The ARC Prize more directly states that economic value is an incorrect measure of intelligence, defining it instead as "a system capable of efficiently acquiring new skills beyond its training data."

The conjunctions are also different.

Anthropic’s Dario Amodei avoids saying AGI, instead referring to it as “powerful AI.” Microsoft’s Mustafa Suleyman coined the term “artificial capable intelligence.”

Meta and OpenAI have recently skipped AGI altogether and are directly discussing "superintelligence."

The schedule is even more disorganized.

In September 2024, Altman said that systems pointing toward AGI are emerging, and even suggested that superintelligence could arrive within a few thousand days.

A year later, GPT-5 was released, and it admitted, "We're still missing something quite important," but didn't specify what it was.

A year later, Brockman said at the Astra launch, "Welcome to the AGI era."

At DeepMind, Shane Legg has held a 50% probability of achieving AGI before 2028 since 1999; however, Demis Hassabis says it will still take five to ten more years. The timeline appears to be shortening, but only because “the definition of AGI is being diluted.”

This is the situation we’re in now.

The same word, defined in four different ways by four different groups.

The model company holds the product narrative, showcasing benchmark scores without making definitive claims;

The chip giant, holding the narrative around computing power, is eager to announce that a new era has arrived;

The benchmark institution holds the test ruler, firmly clinging to the bottom line and refusing to acknowledge it;

Major players like Microsoft hold the contract terms and sit back calmly, calculating the numbers according to the rules.

Over the past four years, the entire industry has competed to be the first to develop AGI.

Recently, the narrative has shifted directly from "another one to three years" or "as soon as six months" to "it has already arrived."

Yet, no one has yet produced a universally accepted set of AGI evaluation criteria.

Reference materials:

https://x.com/JensenHuang/status/2096700264569090384

https://www.bloomberg.com/news/features/2026-09-04/what-is-agi-openai-anthropic-race-for-artificial-general-intelligence

https://www.axios.com/2026/09/03/openai-astra-gpt-6-agi-brockman

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation; edited by Yuan Yu.