On April 27 OpenAI’s Microsoft deal stopped tying its payments to technological progress. On September 3 its president welcomed the “AGI era.” OpenAI’s own test of economic work was not in the launch materials.
I already wrote that Sam Altman gives the same Senate committee opposite testimony depending on his market share, and that the only variable that ever moves is the money. Dario Altman was about the man. This is about the word.
On April 27, 2026, OpenAI and Microsoft rewrote their deal so that OpenAI’s revenue-sharing payments run through 2030 “independent of OpenAI’s technology progress,” instead of ending the day OpenAI was judged to have reached AGI. On September 3, president Greg Brockman closed the GPT-6 Astra briefing with “Welcome to the AGI era,” Axios reported, and, asked about AGI, said there was “no contractual AGI triggering anymore” and called it a “mission concept or spiritual concept.” One hundred and twenty-nine days between the amended deal and the welcome message. The launch materials did not show how Astra meets the charter’s definition of AGI. A launch briefing is not that demonstration.
Read the two dates together. Then read the claim against the definition still in the charter. That is the article. The rest is me showing my work.
The definition is still there
OpenAI’s 2018 charter still defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” You could argue with it. You could also check it, because it pointed at payroll.
The commercial deal added a different test. According to The Information, as reported by TechCrunch, the 2023 Microsoft contract tied AGI to systems that could generate at least $100 billion in profit, a payout threshold rather than a capability one. In August 2025 Altman told CNBC the word was “not a super useful term”. In April 2026 the amended deal cut AGI loose as an off-ramp from OpenAI’s revenue-sharing payments to Microsoft. On September 3 Brockman left the judgment to the reader, then added: “For me personally, I do think we’re there.”
Economically valuable work in the charter. A profit threshold in the reported contract. A personal judgment at the launch. Three different tests, and the story where one cleanly replaced the next was never true. The charter’s words never went anywhere. The same word, walked from a claim you could test to a mood you cannot: the move I traced through Altman’s Senate testimony, run this time on the goal in the charter. What the welcome message never showed is that Astra had met it.
What the scores grade, and what got left out
Astra’s case is numbers, and they are OpenAI’s own. FrontierMath Tier 4, 98%. ARC-AGI-3, 99.9%. ExploitBench, 100%. Take them at face value. A math tier, an exploit set, a puzzle suite: closed problems with answer keys, the domains where a model can be graded like an exam. Astra tops the exam.
One benchmark was missing. GDPval is OpenAI’s own test of economically valuable tasks across 44 occupations, published in September 2025, its language sitting closer to the charter’s than anything else the company runs. OpenAI published plenty at the launch, AutomationBench and BenchCAD among them. The one it did not publish, in the announcement or the system card, was GDPval, the closest thing it has to grading the job rather than the exam. Even GDPval scores one-shot tasks that fall short of whole jobs held over months, and a high number would not on its own prove AGI. Its absence does not disprove AGI either. But leaving your own economic-work test out of a briefing that welcomes the “AGI era” is the omission worth asking about.
None of this is new to anyone who reads this newsletter, and almost five years into the LLM era a benchmark reads as what it is, a number off a test rather than the work itself. I have shown in China Built the Grid that the gap between the best model and a cheap one on real work is small enough that a lead a research lab notices is one a working engineer often cannot feel. The models keep getting better, and some of the published Sol-to-Astra gains are large: AutomationBench moves from 13.3 to 41.4. Progress is real and worth measuring. It does not get to call its own finish line crossed. The architectural objections live in Country of Geniuses, where I read Anthropic’s own system card: after follow-up clarification, none of sixteen technical staff judged Opus 4.6 more likely than not to become a drop-in entry-level researcher within three months of better scaffolding. That was a verdict on one model and one role, not proof of what no future model could do. Here I am testing a launch claim, not settling the architecture.
The benchmark is not the job
A launch briefing is not a field record. Neither is a wall of benchmark scores. We have seen the shape before. When Anthropic shipped Fable 5 in June, Endor Labs tested it two ways: mid-table with Claude Code, at 59.8% functional and 19% security, then top of its fair board at the time with Cursor, at 72.6% and 29%, the same weights in a different harness. Both are real. Neither is the job. Endor was scoring model-and-harness combinations on vulnerability tasks, not months of work inside a client’s system.
I have written this pattern twice already, both times about Anthropic. The percentage was always the pitch; the product was whatever leaked three months later, where readers found a 3,167-line function in print.ts, whatever wrote it. It runs across companies. Selling the story of the lead in place of the lead is how Altman spent his own head start, and Astra is that same shape at a higher number. Saturated scores, a welcome message, and a claim the launch itself cannot establish. That is the tell. I cannot prove it is deliberate, and I do not need to, because the dates carry the weight.
Most aligned. Harder to read.
The guardrails matter more than the word games. OpenAI calls Astra its most aligned model. Its own system card, published the same day, reports “a substantial decrease in chain-of-thought monitorability compared to previous models” in its evaluations. That finding is about chain-of-thought; action-only monitorability improved in the same report. Most aligned, by their measure. Harder to monitor, by their own admission. “Most aligned” is a claim about behavior; reduced monitorability is a limit, in some channels, on how that behavior can be checked, and the first does not cancel the second.
What the calendar says
I will not tell you why they did it. I cannot read intent and I will not dress the timing up as a confession. I can read a calendar.
The December before, OpenAI reportedly called an internal “code red” when Google’s Gemini 3 beat it on benchmarks, the same sort of scoreboard it has now topped and called AGI. It closed a round in March at an $852 billion post-money valuation. In June it confirmed a confidential S-1 submission. Set those beside the two dates at the top. The amended deal came before the launch, and the valuation and the filing sit beside that sequence as context, not as proof of why Brockman chose those words. I took apart the machinery of manufactured permission and manufactured danger before, the same men reaching for the same lever on the same schedule.
Those dates are a sequence, not a motive. The deal changed before the claim. GDPval was not in the launch materials. That is what has dates on it.
Decide for yourself
Brockman says he leaves it to the reader to decide whether Astra qualifies. Fine. Here is the only definition I have use for. AGI understands a domain. It finds the one detail a solution turns on, the small thing that makes a fix hold or brings it down, the way a rivet on an airframe looks trivial until you know what it holds. That is the work: knowing which detail matters, and why. None I have run does it. If yours does, you are ahead of my hundreds of projects, and I want to see it. They assemble answers that read right and miss the detail that decides the fix, and I still have to catch it. That is the test I care about: can I delegate the work and verify the result. A launch score does not answer that. If these models post the performance the numbers claim, why am I still babysitting this shit on every project?
What he handed me is a launch. The “AGI era” is the sticker on the box. The charter still offers a promise you can hold the company to. The welcome message did not show that the model kept it. I will keep writing code with one vendor’s model and reviewing it with another’s, switching the moment the bill or the quality moves, and the day one of them decides this piece earned me a ban, I will look for a Chinese open-weight model I can run on my own hardware. The label on the box will not decide whether I use it. Its work will. It will just do the work, or fail at it, where I can watch which.
Astra might be the best model anyone has shipped. The charter’s definition is still there. The launch did not establish that Astra meets it. Decide for yourself. I have.