The pattern
A launch announcement is written to be shared, quoted and believed immediately. It is not written to be checked. That is not a conspiracy, it is incentives: nobody in the chain between the builders and the press release is rewarded for saying "we are not sure yet." So the announcement runs hot and the correction runs quiet, and most people only ever read the first.
I spent 2026 building a ledger of 132 public claims from fifty named AI leaders, and the gap between the announcement and what came after it is where most of the real information turned out to live. Three cases from the ledger show the shape. The clock runs at a different speed in each one, a day, a quarter, twenty-one months, but the pattern is the same: a triumphant statement made when it was cheap to make, followed by a correction that arrived only when keeping the statement got expensive.
Case one: GPT-6 Astra, twenty-four hours
On September 3, 2026, Sam Altman told the world that GPT-6 Astra marked the arrival of "the AGI era." Twenty-four hours later he was on X apologizing for a "messy rollout." Paying ChatGPT Plus and Pro subscribers had been locked out while early access went to a narrow cybersecurity testing program. OpenAI offered a banked daily reset as compensation and, within 48 hours, extended access to Pro, Enterprise and Business Premium tiers.
The independent evidence arrived in the same week. Artificial Analysis scored Astra 67 on its Coding Agent Index, tied with Claude's Opus 5 and five points behind Anthropic's Fable 5.1, which had shipped two days earlier. On the separate Intelligence Index, Astra scored 61 against Fable 5.1's 66. Then, three days after launch, OpenAI's own chief scientist, Jakub Pachocki, published an essay called "An Alien Mind," arguing that no lab, including his own, has solved alignment well enough to justify scaling at full speed. That is the company's own top research officer, days after his CEO declared a new era.
To be fair, Astra was the first OpenAI model to reach the "Critical" threshold on the company's own cybersecurity Preparedness Framework, a real engineering result whatever you think of the framing. A model can be a genuine advance and still not be what its launch language claimed.
One more detail. In June 2025, Altman wrote that "intelligence too cheap to meter is well within grasp." Astra launched at $10 per million input tokens and $50 per million output tokens. You can argue he meant relative to an earlier cost curve, but notice that the argument moves the claim from a number you can check to a comparison you can only argue about. That move is part of the pattern too.
Case two: Meta's open-weights pledge, twenty-one months
On July 23, 2024, Mark Zuckerberg published "Open Source AI Is the Path Forward" and wrote, "I believe that open source is necessary for a positive AI future." On April 8, 2026, Meta Superintelligence Labs shipped Muse Spark, its first frontier model on an entirely new stack. It shipped closed.
The correction came in stages, and none of them named the reversal. Alexandr Wang, Meta's Chief AI Officer, called openness "a deployment decision, not a company identity." In a May 2026 podcast appearance he said Muse Spark had "triggered some high risk areas in the course of early training, particularly around bio risk," and Meta's own safety report, submitted to arXiv in May 2026 with 119 authors, assessed the model's chemical and biological risk category as likely reaching what the company calls "high risk" before mitigations. Credit to Wang: that is a specific, checkable reason.
Then, on August 10, 2026, Zuckerberg published a second essay, "The Future is for Everyone," roughly 6,500 words arguing that the real danger in AI is concentration of control, not capability. It makes the case for openness in the abstract, four months after the flagship shipped closed, and it does not mention that contradiction anywhere in its own text. The same day, Meta released Muse Glimmer, a thirty-billion-parameter model distilled from Muse Spark, under a genuinely open Apache 2.0 license. A real partial reopening. Not the flagship. On September 8, Meta launched Muse, a consumer agent built on exactly the philosophy the essay described.
The money explains the sequence better than the essay does. On April 29, 2026, Meta disclosed AI spending of up to $145 billion for the year and, asked about the return, Zuckerberg reportedly called it "a very technical question." On July 29, second-quarter free cash flow had collapsed 91 percent, to $784 million from $8.55 billion a year earlier. April: enormous spend. July: a cash warning. August: a philosophical reframing. September: a product that needed the reframing to land. I cannot see inside Zuckerberg's head. I can see the sequence.
Case three: Grok 5, a missed quarter
This one needs no interpretation. At a Baron Capital investor conference in November 2025, Elon Musk said Grok 5 would ship in the first quarter of 2026, with roughly six trillion parameters, and put its odds of achieving AGI at about 10 percent. The window passed. As of September 10, 2026, the day I closed the ledger, Grok 5 had not shipped. The AGI-probability half of the claim is unfalsifiable, since no criterion for resolving a "10 percent chance" was ever offered. The shipping half is refuted, on the calendar.
The backdrop matters. Musk's year otherwise reads as a run of wins: xAI folded into SpaceX in February 2026, the combined company went public in June at a valuation near $1.77 trillion, and it closed a $60 billion acquisition of Cursor in August. A missed ship date is small against that. But alongside it sits sworn testimony from April 30, 2026, in the Musk v. OpenAI trial, that xAI had "partly" trained Grok on OpenAI's outputs, and by late July the newly public stock had fallen roughly 50 percent from its post-IPO peak. Public claims running ahead of what the product and the legal record could support. That is the pattern, at a different speed.
What connects the three
In each case the triumphant statement was made when it cost nothing but words and bought real goodwill. Zuckerberg's 2024 essay generated enormous positive press. Musk's Q1 date generated excitement months before a product had to exist. Altman's "AGI era" line generated the biggest headline of the year. And in each case the walkback arrived only once circumstances changed enough that keeping the statement became expensive: a model that failed its own bio-risk review, a timeline that slipped, a rollout that locked out the people paying for it. A pledge made when it is cheap is not yet evidence of anything. The evidence arrives later, when keeping it starts to cost something.
Five checks before you believe the launch post
This is the Launch Announcement Checklist from the book, cut to five questions. Run them before you buy, switch vendors, or tell your team to adopt something.
1. Who is making the claim, and do they have a direct financial stake in you believing it? Every launch has a stake in sounding as big as possible. That does not make it false. It means you weight it.
2. Is the number self-reported or independently measured? Astra's "99.99% success on instruction hierarchy tests" comes from OpenAI's own system card. Artificial Analysis's 67 on the Coding Agent Index comes from an outside evaluator with nothing to sell. Those are different kinds of evidence, even when they sound equally confident.
3. Has anyone outside the company tried to reproduce or confirm it? For Astra, yes, within days, and the numbers came in lower than the framing. For most launches the honest answer on day one is "not yet," which should move your confidence down, not up.
4. What happened in the first 48 hours, not the first 24? Altman's apology came within a day. The real state of a product surfaces when real customers use it under load, not in the demo.
5. Does the claim have a fixed, checkable resolution, or is it worded so it can never be proven wrong? "The AGI era has arrived" resolves to nothing in particular. "Ships in Q1 2026" resolves to a calendar, which is why it could be scored.
If two or more answers land on self-reported, unconfirmed, or no fixed resolution, treat the launch as marketing until the 48-hour picture and outside evaluators fill in the rest.
Where these cases come from
All three are worked in full in The Claims Ledger, the book I wrote from a year of scoring what powerful people in AI said against what happened next: 132 claims, fifty people, window closed September 10, 2026, full ledger in the back. It is $19, PDF and EPUB, at /book. The five checks above are free and complete without it.