The Take — Calling Astra AGI is flippant. The field's silence is worse.
M.G. Siegler is right that calling GPT-6 Astra "AGI" is marketing. He is wrong that OpenAI alone is to blame. When Greg Brockman stood in front of reporters last Thursday and told them "we are now in the AGI era," he was performing exactly the move Siegler calls flippant — but he was performing it on a stage the rest of the frontier-AI field had abandoned to him. Nobody serious has bothered to define the term in a decade, and the bill is now due.
I think the AGI argument is the wrong argument. The interesting one is who is allowed to say the word and get away with it, and what that tells us about the rest of the labs.
What Brockman actually said, and why it's more honest than it sounds
The quote that's drawn the fire is the one that reads like a personal announcement: "If we fast-forward a couple of years, and we look back and say, 'When was it, really, that AGI was created?' I think it's going to be about this time, and I think it might be about this model." That is flippant if you read it as a scientific claim. It is not flippant if you read it as Brockman made it plain in his Stratechery interview the same week, where he described AGI as having drifted from a contractual obligation to a "mission concept or spiritual concept" and refused to commit to a definition. Our deep dive on that interview makes the move more visible: he is saying the word, then conceding he cannot say what the word means. That is not deception. It is closer to an honest admission that the term has been a vibes-based marketing asset since Sam Altman co-wrote a blog post called "Planning for AGI and Beyond" in 2018.
The launch context is where it starts to land. As we reported Thursday, OpenAI also published numbers that are, by any honest read, the strongest capability claim a frontier lab has shipped. Astra hit 97.6 percent on FrontierMath Tier 4 v2, 100 percent on ExploitBench, and 99.9 percent on ARC-AGI-3, all on more than 100,000 GPUs at the Stargate site. ARC Prize's Greg Kamradt independently confirmed the ARC-AGI-3 number, though with the same model scoring 62.7 percent on the standard harness — a 37-point gap between the public number and the canonical one that nobody at OpenAI has yet explained. Artificial Analysis put Astra at 61 on its Intelligence Index, five behind Anthropic's Claude Fable 5.1. The benchmark brief that night called this exactly what it was: a model measurably better at coding and hallucinating less, but trailing on general intelligence, with a 2.5x price hike attached.
So the launch was a real capability launch wrapped in a word that nobody has earned the right to use, marketed by a company that has spent the year making that word mean whatever it needed to mean in the moment. Siegler's critique nails that. The interesting question is why the field lets one company own the term.
The counter-case, stated fairly
The strongest version of OpenAI's position is that AGI was always a marketing word, and Brockman's quoted line is the first time a frontier-lab executive has admitted that on the record. The original OpenAI charter used the phrase "highly autonomous systems that outperform humans at most economically valuable work" — a definition with the charm of being specific enough to argue with and vague enough to redefine every quarter. Microsoft's contractual version, never public in full, leaked the phrase "general intelligence at human level" and was the basis of a multi-year fight over profit participation that ended when Microsoft agreed to a structure that does not require AGI to ever be declared at all. By the time Brockman sat down with Thompson, AGI had been loosened from a contract term to a "spiritual concept." That is closer to a confession than a claim, and you could read it as the most candid thing OpenAI has said about the term in eight years.
There is also a defensible position from the benchmarkers. ARC Prize's Kamradt designed ARC-AGI-3 to be the benchmark that AGI claims would have to clear, and OpenAI clearing it (on a non-standard harness) is at least an empirical event rather than a vibes event. DeepMind published its own framework last year — "Measuring progress toward AGI: a cognitive framework" — that explicitly rejected the binary in favor of capabilities grouped into cognitive categories, which is the only careful position any frontier lab has taken on the term. Anthropic, by contrast, has avoided the word almost entirely in product launches and lets its capability disclosures speak. The result is that the only lab talking loudly about AGI is the one with the strongest commercial reason to talk loudly about AGI, and the labs with the most credibility on the question have chosen silence.
Why the take still holds
Two things keep the take honest even after you grant Brockman the benefit of the doubt.
The first is that the AGI label is doing real work for OpenAI that the company has not earned the right to claim. The Microsoft restructuring, the $40 billion revenue run rate, the staged rollout to paying customers while water utilities and local governments get first access — each of those is defensible on its own merits. None of them is helped by an announcement that the company has crossed a civilizational threshold, and the announcement has nothing to do with the technical reality. Brockman's "spiritual concept" line is honest; the marketing around it is not. A company that says "we have crossed into AGI" on Thursday and "we cannot define AGI" on Friday has not earned the public trust that the second statement asks the reader to extend. That is the flippance Siegler identified, and it is real.
The second is that the rest of the field has had a decade to define the term and chose not to. DeepMind's cognitive framework is the only serious attempt; everyone else has either avoided the question or repeated OpenAI's marketing copy. That is the part of the problem OpenAI did not create alone. A term without a definition is a term anyone can use. The labs that complain when OpenAI uses it should have published their own operational definition years ago and let the benchmarkers grade it. None of them did. The 2018 OpenAI charter's "outperform humans at most economically valuable work" is the closest thing to a working definition the field has, and it has been quietly dropped from the company's own communication. That is not a one-company failure. It is a field failure, and Siegler's critique is miscast as long as the response is to dunk on OpenAI rather than to ask why the term is up for grabs at all.
There is also a cost-of-silence problem nobody has named. The longer the field refuses to define AGI, the more the term accrues to whoever is loudest. Right now that is OpenAI, because OpenAI has the most to gain from loud. Anthropic's deliberate avoidance reads as prudence, but it also means that when regulators ask "what is this thing," the only public answer is the one OpenAI gave them. The Massachusetts state-AI bill story — Anthropic backing state-level rules while OpenAI pushes a slower national alternative, as our launch brief reported — is a small version of the same pattern: Anthropic has the more defensible position but a quieter press operation, and OpenAI's counter-position is the slower and weaker option that is more likely to win for the simple reason that it is louder.
What would change my mind
Three things. First, if a frontier lab outside OpenAI publishes an operational definition of AGI with named capabilities and benchmarks, and the field agrees to it. Right now DeepMind's cognitive framework is the only serious attempt and it has not been adopted. If Anthropic, Meta, and Mistral sign on to a shared definition and refuse to use the word outside it, the term stops being marketing and starts being a target. Second, if OpenAI publishes the harness difference between its 99.9 percent ARC-AGI-3 number and ARC Prize's 62.7 percent standard-harness number. The gap between those two numbers is the gap between a capability claim and a marketing claim, and resolving it would either validate the launch rhetoric or expose it. Third, if a regulator writes the term into binding law. Once AGI is a legal definition rather than a vibes-based marketing asset, the marketing stops working, and so does the flippance.
Until then, the right read is that Brockman is being more honest than his critics want to admit — he is saying the word is meaningless and using it anyway, in that order — but the rest of the field is the bigger problem, because they had the chance to do the careful definitional work and they let OpenAI own the term by default. Siegler is right that this is marketing. He is wrong that one company is the source of the marketing. The source is a decade of silence from everyone else.
A word that no one else is willing to define is a word the loudest speaker gets to keep. Is that the field's failure or OpenAI's win? Tell us in the comments.
