
Anthropic told investors on 10 November that the alarm it built for runaway AI progress has gone off, and it is still lined up to list on Nasdaq this week. The twist tripped it: the company says it made that call based on what it expects from future models, not on the ones you can use today.
Most of us have felt a small version of this at our desks, a task that ate an afternoon last year and now takes twenty minutes. Anthropic is describing the same shift inside the team that builds the models, and it has put the claim into a document investors will rely on.
I read the three documents that matter: METR’s audit of Anthropic, the AI Futures Project’s new forecast, and the filing itself. I do not think this is marketing; I do not think it is the end of the world either. What I found was narrower and more uncomfortable. Much of the evidence behind the rate is redacted, so the public is being asked to trust a number it cannot inspect.
One filing, three versions, 15 days
The cleanest way to see how fast this moved is to read the filing’s wording as it changed. The comparison comes from FourWeekMBA’s side-by-side of the three versions, and the filing itself sits on the SEC’s site.
On 26 October, the original S-1 said that Anthropic “may not identify a significant acceleration in our rate of development until after it has occurred.” That is a company telling investors it might not notice the thing until it is over.
On 2 November, Amendment №1 said that Anthropic “may imminently meet” the definition in its own policy, based on “internal estimates and our expectations regarding future models, rather than our current models.” It also said it had designated no current model as past the line.
On 10 November, Amendment №2 said Anthropic had “made a determination that we have met this threshold.” The basis stayed the same, what it expects from future models, and the sentence about current models was gone. In the same amendment, the risk factor about models behaving badly moved from ninth place to first.
So the alarm went from “we might miss it” to “we may be there” to “we are there” in 15 days, and at no step did the justification become a measurement of a released model. I do not read that as a gotcha. A forward-looking test is a defensible way to write an alarm because you want it to ring before the fire. It means this is a forecast Anthropic made about itself, and what crossing the line obliges the company to do is set by its own policy, not by an outside rule.
An outside audit says the clock runs at about double speed
The filing reads this way because of a report METR published on 5 November. Its staff spent 11 weeks, from 12 August to 29 October, with access to benchmarks on unreleased models, training metrics, employee surveys and internal analyses. The finding is one sentence:
“Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025.”
Read the qualifiers before the number. The central estimate is 1.9x, with a 95 percent range of 1.5 to 2.8x, and it is about capability gains on well-specified tasks, not about AI getting twice as smart at everything. The second number is the one that carries the story: “a majority (perhaps ≈80%)” of the current acceleration, METR guesses, is because of AI. METR itself marks that as highly uncertain, and says non-AI factors probably drove much of the late-2025 speed-up.
What METR says it is confident about is shorter. The acceleration was not driven by faster growth in training compute. AI is not yet making fully autonomous contributions to the most important parts of model training, though METR sees “signs of life” that it might soon. And METR did not find any clear bottleneck for the next 12 months, with low confidence, because its own view changed several times during the audit.
Then comes the part I keep returning to. Anthropic redacted most of the evidence METR found most informative, which Anthropic described as removing confidential information that could speed up other developers. METR says the public must rely on its judgment “much more than would be ideal,” as The Decoder reports. The same coverage adds that METR thinks the threshold may have been crossed as early as 15 July, that the figure implied by Anthropic’s internal August risk report sits at 1.97, a hair under 2, and that the true factor could be anywhere from 1.3 to 3.
A hair under 2, judged by an auditor who cannot show you its best evidence. That is the quality of the thing the market is pricing.
The forecasters moved 19 months in under three
The AI Futures Project, the group behind the AI 2027 scenario, published its update on 10 November under a title that tells you where it landed:
“Looks Like the Intelligence Explosion Will Happen in 2027 After All.”
In February, it had graded its own scenario at about 65 percent of the pace it predicted. This is a reversal, not a continuation.
Their arithmetic is easy to follow. Take METR’s roughly 2x, credit AI with about 80 percent of it, and you get an AI-driven acceleration of about 1.7x. Drop that into their own scenario and 1.7x arrives around April 2027, about five months ahead of where they had it.
The individual medians moved more. Daniel Kokotajlo’s date for a Superhuman AI Researcher, their milestone for an AI that out-researches the best humans, went from December 2028 in mid-August to May 2027. Eli Lifland now sits at August 2027, and none of the medians the authors state land later than November 2027. They also put the chance of a Top-Expert-Dominating AI, one that could automate all white-collar work, at between 35 and 65 percent before 2028. Every one of those numbers assumes nobody deliberately slows down, and the authors warn that their modelling may contain errors.
Here is my problem with treating any of it as a date. A forecast that moves 19 months in under three is a measure of uncertainty, not of knowledge. And this one stands on a guess (the 80 percent), which stands on evidence METR says is largely redacted. I would take the direction seriously and the month with a very large grain of salt.
Anthropic says Claude now leads a quarter of its own research
The company has its own gauge, and it is more specific than most people realise. In its research on measuring the pace of AI development, Anthropic says Claude “leads” 26 percent of its AI R&D work as of August 2026, up from under 1 percent in February.
“Leads” has a precise meaning. On Epoch AI’s automation scale, it is level four: the AI can complete most of a task end to end from a high-level prompt while a human supervises. Over 90 percent of the work sits at “AI collaborates” or higher, and no measured area is fully autonomous. So a quarter does not mean that Claude is doing a quarter of the research alone.
It means that a quarter of the work is now supervised rather than done directly.
The method deserves the same skepticism I would give any company grading itself. Anthropic sampled 20 percent of staff in each model’s R&D department every week in July. A Claude agent mapped their tasks from Slack and internal documents, and a separate Claude judge rated how automated each task was. The judge matched staff ratings 59 percent of the time, while staff agreed with each other only 35 percent of the time. Anthropic concedes the judge could make “the same kinds of errors as the model it is checking.”
The anecdote that travelled furthest comes from the roadshow Anthropic kicked off for its IPO. In its roadshow video, the company says Claude Mythos got Claude running on AMD chips “in just three days, with basically no human intervention,” and co-founder Tom Brown says Claude “ported or wrote almost 100% of the kernels.” OfficeChai’s report on the roadshow frames the comparison as dozens of engineers and months of work for the same job, and adds, as reported, that an Anthropic engineer connected the model to the hardware and told it to start.


