Is GPT-6 Astra Really AGI? OpenAI Can’t Even Agree With Itself
In a nutshell
everything on the web starts with the domain
Yesterday we urged caution on OpenAI's claim that its new GPT-6 Astra model marks the arrival of artificial general intelligence. One day later, the strongest argument for that caution is coming from inside OpenAI itself. The company that declared the AGI era cannot, it turns out, agree internally on whether the AGI era has actually begun — and the independent numbers are complicating the story further. This is not us versus OpenAI. This is OpenAI versus OpenAI.
The Claim That Split Its Own Maker
Look closely at who said what, because the cracks are right there in the record. The words "Welcome to the AGI era" came from OpenAI's president, Greg Brockman, spoken aloud in a press briefing and explicitly framed as his personal view — "I do leave it up to the reader to decide for themselves; for me personally, I think we're there." He even conceded the term is "fuzzy." Now set that beside two other facts. OpenAI's own written announcement of Astra pointedly avoids the word AGI altogether, calling the model only its "most intelligent and aligned" system. And OpenAI's chief executive, Sam Altman, has said the company is "not quite" at AGI — having dismissed the very term, not long before, as a nearly meaningless marketing label. So within a single company, on a single launch, the president says AGI has arrived, the CEO says not quite, and the official text refuses to use the word. When the people who built the system cannot agree on what they built, that disagreement is the story.
What the Independent Numbers Say
The outside evidence adds a second layer of caution, and it deserves to be stated carefully and fairly. Astra is, by every serious account, a genuine leap in capability — its autonomous computer use and software-engineering results are real, corroborated by early testers, and impressive on their own terms. But the picture of total dominance did not survive contact with independent benchmarks. The evaluation firm Artificial Analysis scored Astra level with OpenAI's own older model on its intelligence index, and behind several rival systems, at a substantially higher cost per task. Another independent evaluator recorded a real, record-setting jump in capability while cautioning that genuine AGI would sit far above the current trend line, not merely at a new high point on it. The takeaway is not that Astra is weak — it plainly isn't. It is that "most capable model yet" and "artificial general intelligence" are two very different claims, and the data supports the first far more than the second.
The Deeper Problem: A Word With No Meaning
Underneath all of this sits the flaw that makes the whole debate slippery: there is no agreed definition of AGI. OpenAI is measuring itself against a bar it wrote, and even its own leaders describe that bar as fuzzy. When a term has no shared, testable meaning, declaring you've reached it is less a scientific finding than a choice of words — and words can be chosen for many reasons, including the fact that the largest IPO in history is weeks away. This is why the internal disagreement matters so much. It is not a small communications slip. It is proof that "AGI" is not a line the model crossed, but a label people are choosing, or refusing, to apply. A milestone that some of its own creators recognise and others do not is, by definition, not a milestone. It is an argument.
Claim and Counter-Claim
In fairness to the AGI reading: something real did happen this week. A system that can autonomously operate a computer, write and fix software end to end, and reason across enormous context is doing things that genuinely blur the line between tool and agent, and reasonable people looking at Astra can feel the ground shifting beneath the old definitions. Brockman may simply be first to say out loud what others will say in a year.
But the counter-case is now stronger than it was yesterday, because it no longer relies on outside skeptics — it relies on OpenAI's own executives and on independent data. When the CEO hedges, the written announcement demurs, the chief scientist warns that alignment has not kept pace and that the model is harder to monitor, and independent benchmarks decline to crown it, the honest conclusion is not "AGI has arrived." It is "a very capable model has arrived, and a very loud word has been attached to it." Our position, unchanged from yesterday and reinforced today: be impressed by the machine, and be patient with the label.
The European Perspective
For Europe, the past twenty-four hours sharpen the argument we made when Astra launched, and turn it from principle into evidence. Yesterday's point was structural: one company can declare a civilisational threshold with no independent body able to check it before the fact. Today's development shows why that gap is so dangerous — because the claim turns out to be contested even inside the company making it, and only independent evaluators, doing the patient work of measurement, began to put it in perspective. That is precisely the function Europe's regulatory architecture is meant to serve, and precisely the capacity it does not yet have at sufficient strength.
The lesson is not that OpenAI lied; it is that "AGI" is becoming a word with real regulatory and contractual weight — triggering safety obligations, shaping public policy, moving markets — while remaining scientifically undefined and, apparently, internally disputed. If a term this consequential can be declared true by a president, doubted by a CEO, and omitted by the official text all at once, then leaving its definition and verification entirely to the companies that profit from it is not a tenable position for any society.
Europe cannot build Astra. But the one thing this week proves it must build is the capacity to independently answer the question OpenAI cannot answer for itself: not "is it impressive," but "is it what you say it is." Until someone outside the labs can answer that, every future AGI claim will be exactly what this one is — a word we are asked to take on faith.
We are not first. We are right.
SOURCES
Al Jazeera — OpenAI unveils GPT-6 Astra amid rising scrutiny and safety concerns
Time — OpenAI Launches GPT-6 Astra Amid AGI Claims and Mixed Benchmark Results
VentureBeat — "Welcome to the AGI era": OpenAI launches GPT-6 Astra
The New Stack — OpenAI launches GPT-6 Astra and says welcome to the "AGI era"