A note before we start. This piece is about Anthropic, the company that makes Claude, the AI system that helps us write. We are Humanity and AI, a civic technology and research group in Oklahoma City. We applied to Anthropic’s Fellows research program this summer, an application now on hold until a future round. We report on Anthropic the way we would report on any company, and we state the weak points plainly. The full disclosure is at the end.
When a company wants to sell shares to the public, it files a long document with the government that lists, by law, the serious risks a buyer should know about. The document is called an S-1. The company files it with the Securities and Exchange Commission, the federal agency that polices stock markets, and most of it is dry. One company’s risk list is not.
On November 2, Anthropic, the maker of the Claude AI models, filed an updated version of that document ahead of its planned stock-market debut. Its risk section said Anthropic “may imminently” reach a line it drew for itself: the point of “dramatically accelerated progress attributable to AI.” That is the moment when AI starts doing enough of the work of building better AI to speed progress up dramatically. “May imminently” is a careful legal way of saying soon. The company is clear that this is based on what it expects from future models, rather than on its current ones. An earlier version of the company’s safety policy, the public rulebook Anthropic writes for itself on handling dangerous AI abilities, put a number on it: two years of progress compressed into one, roughly a doubling of the pace. The company revised that threshold in July. The same paragraph adds that the company has “in the past not identified” such a speedup “until after it has occurred.”
Eight days later, on November 10, Anthropic filed again. This time it called the line its “AI research and development capability threshold,” the part of its safety policy about AI speeding up its own research, and the sentence changed: “We made a determination that we have met this threshold.” The company says it based that call on “what we expect from future models,” not on what it has seen its current models do, though it adds that its use of AI “has already been accelerating our research and development.” It says it is still studying the underlying data and will publish further reports. The admission about missing past speedups stayed in. In plain words: by its own measure, Anthropic now counts itself over the line, on the strength of what it expects its next models to do.
Those facts are the story. A frontier AI company, meaning one of the handful of firms building the most advanced AI systems, told its future investors, in a legal document where misleading them carries real legal consequences, that the moment AI begins to accelerate AI has arrived by its own measure, and that it has a track record of only seeing these jumps in the rear-view mirror.
What the claim is, and what it is not
The headline version of this (“AI is about to build itself”) runs far past what the filing actually says.
The filing does not say any Claude model in use today has crossed that line. The company’s determination rests on what it expects from models it has not yet released. That is a judgment about where things are headed, made by the company itself, about its own work.
There is now an outside measurement to set beside the filing. METR, a nonprofit that runs independent tests on AI models, spent eleven weeks working inside Anthropic, from August 12 to October 29, with access to internal data and to staff. On November 5 it published what it found: Anthropic is likely improving its models at “around twice the rate” it did between March 2024 and October 2025, on what METR calls well-specified tasks, meaning work with a clear goal that can be checked. In plain words, the pace has roughly doubled. METR says it would not be surprised if the true figure were closer to 1.3 times or 3 times, and it guesses that most of the speedup, perhaps 80 percent, comes from Anthropic using AI in its own research.
Three cautions come with that number, and METR states all of them. First, the speedup comes mainly from AI helping researchers, not from AI doing research by itself; METR says fully independent AI contributions are not yet a large part of progress, though they may be soon. Second, METR did not judge whether Anthropic has crossed the line in its own policy; it says it did not focus on the company’s thresholds.
The third caution is the sharpest. Anthropic had the right to strike private material from the report before it was published, and it “redacted a majority of the evidence we found most informative,” meaning it blacked that evidence out before the public saw it. METR’s own conclusion is that “the public has to rely on METR’s judgment much more than would be ideal.” In September, inside its review of a single model, METR had quoted a rougher estimate from a separate METR team, about 1.5 times, which it did not vouch for. That number measured something different, the speedup against a world with no AI help, so the two are not one figure revised. We rely on the November report because it is METR’s own and far more thorough.
So the honest shape of it: a serious company, in a serious document, says it has reached its own line, based on what it expects next, and the closest outside look anyone has had says the pace has already roughly doubled, mostly because of AI, while the people doing the work say the most important parts are not yet done by machines on their own. Both things are true. Neither cancels the other.
Why a company might only see it afterward
The filing admits that Anthropic has missed these speedups before, noticing them only after they happened. The company gives the reason in the same paragraph: much of its evidence comes from “lagging indicators,” signs that only show up after a change has already taken place. This is a property of the thing itself, not a sign of carelessness.
Some systems cannot be predicted from the outside. The only way to know what they will do is to run them and watch. A video from the channel Emergent Garden explains an idea the scientist Stephen Wolfram named “computational irreducibility”: there is no shortcut formula, no peek at the answer, you have to let it play out. Weather is a little like this. So, apparently, is the pace of AI research. If progress can jump without warning, then even the people closest to it can be looking right at it and not see the turn until they are already around it.
The filing is not promising a prediction. It is admitting the limits of prediction, in writing, to the people about to buy the stock.
What we think should happen
We do not pretend to be neutral here, and our rule is to say what we think after we show the evidence. Two things.
First, the measuring should not be left to the companies alone. The best number in this story came from an outside group, METR, after eleven weeks inside the company, and even then the company removed most of the evidence METR found most telling before the public saw it. If an AI company’s own paperwork says it cannot reliably see these jumps coming, then independent, outside measurement, done by people with no stake in the share price, is the only instrument pointed at the road. The filing itself says Anthropic is considering rules that would require human review and approval of some research decisions as its models take on more of that work. That is the right instinct, and the reviewers should include people outside the company.
Second, this belongs in public view, not buried in a legal filing. Whether AI is starting to accelerate AI is not a question for investors and insiders. It touches everyone’s work and everyone’s future. The people it will touch deserve to watch the gauge. That is the whole reason we exist.
Hold us to it
Here is our dated check-in, so this is a claim you can hold us to and not a vague worry. By February 1, 2027, we will report back on two questions, with sources. Has Anthropic published the further reports its filing promises, with evidence that people outside the company can check? And has any independent outside group, with no financial stake, measured an acceleration and been able to show the public its evidence? METR’s November report measured one, and could not show most of the evidence it found most informative. We will tell you plainly what we find, including if the answer to both is still no.
The filing now says the company has reached the turn, by its own measure, and that it has missed turns like this before. The right response to that is not panic and not a shrug. It is to point a trustworthy instrument at the road, in the open, and read it out loud.
Disclosure
We hold ourselves to the standard we ask of others. Our work is produced with substantial help from Claude, an AI system made by Anthropic. This piece reports on Anthropic’s amended filings to sell stock and on two outside assessments by METR, one of a single model and one of the company’s pace of research. Claude drafted this piece, and separate Claude sessions, including Opus 4.8 and Opus 4.6, checked its facts, its plain language and its voice. In July 2026 Humanity and AI, the Oklahoma City organization that publishes this piece, applied to Anthropic’s Fellows research program. That application is on hold until a future round. This piece also names METR in ordinary factual reporting. Nothing in it was shown to, sponsored by, or reviewed by any company or group named in it.
Sources
- Anthropic PBC, Form S-1/A (Amendment No. 1), filed November 2, 2026
- Anthropic PBC, Form S-1/A (Amendment No. 2), filed November 10, 2026
- Anthropic, Responsible Scaling Policy, version 3.4, effective July 8, 2026, with version history
- METR, “Independent Assessment of AI R&D Acceleration at Anthropic in October 2026,” November 5, 2026
- METR, “Summary of METR’s predeployment evaluation of Claude Opus 5.5,” September 22, 2026
- Emergent Garden, “Emergent Complexity” (video)
- Stephen Wolfram, A New Kind of Science, on computational irreducibility