A machine that builds the next machine, and a report that landed before the roadshow

METR published its assessment of how fast Anthropic is improving its own models on Thursday 5 November, and I read the thread at quarter to three on Saturday morning, which is about when I read most things. A small team spent eleven weeks with Anthropic's internal data, from 12 August to 29 October, and came out saying Anthropic is likely improving the capabilities of its models on well-specified tasks at around twice the rate it managed between March 2024 and October 2025, or in METR's words "roughly two years of historical progress over the last year" ^[1]. It guesses that a majority of that speed-up, perhaps 80%, is due to AI, and it didn't find anything in algorithms, compute or data that would slow it down over the next twelve months ^[1].

Anthropic's prospectus, public since 26 October, makes the same claim in sales language, that "each new version of a Claude model directly accelerates development of the next" ^[2], so the independent report agrees with the central pitch of the document Anthropic is using to sell its shares, and it came out in the ten days between that document going public and the earliest date the roadshow can start. I don't think anyone has to have done anything wrong for that to deserve a closer look.

What the numbers say

The main number comes from Anthropic's version of the Epoch Capabilities Index, a benchmark score tracked across its models over time, and since October 2025 it has grown at 1.9 times the earlier rate, with a 95% interval of 1.5 to 2.8 and robustness checks that gave anything from 1.1 to 2.9 ^[1]. Plot every frontier model instead of only Anthropic's and the speed-up drops to about 1.4 times ^[1], so part of the story is Anthropic catching up after a slow stretch, which METR lists itself as a reason it might be overestimating (Anthropic's growth rate in 2024 and 2025 may have been unusually low) ^[1].

METR is confident the speed-up didn't come from growing training compute faster, though the evidence for that is redacted ^[1], which leaves better algorithms, better data and the people, and the people say Claude is doing a lot of the work. In a series of Slack polls Anthropic staff put their own speed-up from the most capable internal model at about 12 times on average, which METR treats as only very loose evidence ^[1], and the gap between the 12 times people feel and the 2 times METR measured is what I'd expect if the model does the parts that are easy to check very fast and leaves the rest of the work about where it was, and METR says something close to that, that the acceleration might be limited to well-specified tasks because most of its evidence measures tasks that are "easy to grade automatically" ^[1]. On 3 November I wrote that the models give us more work than we can check, and this looks like the same thing seen from inside a lab.

Does it accelerate, or not

I think it does, and I think it's a smaller thing than "AI is building itself". Twice the old rate on tasks a computer can grade is a lot, it compounds, and METR's third conclusion is that with no bottleneck in sight there's "a significant risk that further capabilities progress could lead to even faster acceleration" ^[1]. METR also says its view of the bottlenecks changed several times during the work and that it isn't confident in where it landed ^[1], and Anthropic's own filing admits that its evidence on how fast its models are accelerating comes from lagging indicators, so it might not notice a jump until after it has happened ^[2].

Where I'd push back is on reading 2x as the speed of research as a whole. A benchmark that grades itself is the part of the job a model is best at, and the staff told METR that the most important contributions to AI research aren't fully automated yet ^[1]. The parts that are hard to check, whether an idea was the right one to chase and whether a result holds up outside the test that produced it, are the same parts that were slow before, and if they stay slow, twice as fast on the graded half of the work comes out a lot less than twice as fast overall.

The timing

Anthropic filed its draft S-1 confidentially on 1 June ^[3], more than two months before METR started, so the assessment wasn't arranged around a filing nobody knew about, and the dates after that line up closely with the IPO.

Anthropic had the right to redact any non-public information, every piece of information in the public report was approved for release by Anthropic, and METR says Anthropic redacted a majority of the evidence it found most informative for its conclusions, which Anthropic says was to remove confidential information that could help other AI developers ^[1]. The contract also included checking how Anthropic splits its compute between capabilities and safety, and Anthropic changed the scope midway to drop that part before METR reached any conclusion, saying it wanted to keep the work on risk assessment and would come back to compute later ^[1].

I think the boring explanation covers most of it, since a company in registration has its lawyers reading every public word about it, and redaction guidance getting tighter in the days after the S-1 went public is what you'd expect, and METR says the work started in August, it took no payment from Anthropic, and it thinks it "extremely unlikely Anthropic was actively trying to mislead us" ^[1]. METR wrote this as a risk assessment, and the risk is right there in its conclusions for anyone who reads past the 2x.

What bothers me is what reaches the public and when. Someone reading about Anthropic the week before the roadshow gets an independent organisation saying the company's models now help build the next ones at twice the old speed, which is the prospectus's own pitch, and they don't get most of the evidence behind it, because the company selling the shares decided which parts could be published, and the one part that would have said how much compute goes to safety was cancelled before it said anything. Nobody needs bad intent for that to be a problem, since the company being assessed held the redaction pen during its own IPO, and an investor and a safety researcher are now reading the same headline for opposite reasons.

What would change my mind

Anthropic could publish the redacted evidence once the quiet period is over, or enough of it that someone outside METR can check the 2x and the 80% for themselves, and it could put the compute allocation check back in with a date attached. METR's own report says the public has to rely more on METR's judgement than would be ideal ^[1], and I'd like that to be temporary rather than the way these assessments work from now on.

Where I stand

I'm a paying customer, the review panel that grades listings on this site runs on Claude, which is why Anthropic's own listings carry a label, and I wrote about what the S-1 says about agents on the day it was filed, so read all of the above with that in mind. I think the acceleration is real and narrower than it sounds. The roadshow can start on Tuesday, and I'd like to see some of that evidence before then, though I don't expect to.

References

Read on 7 November 2026. Quotes are as published.

1. [1]METR, 5 November 2026. Independent assessment of AI R&D acceleration at Anthropic. Source for the twice-the-rate and two-years-in-one conclusions, the ≈80% attributed to AI, the AECI growth of 1.9x (95% interval 1.5 to 2.8, robustness checks 1.1 to 2.9), the 1.4x across all frontier models, the compute finding, the ≈12x Slack-poll self-reports, the well-specified tasks caveat, the staff view that the most important contributions aren't fully automated, the bottleneck conclusions, the timeline from 12 August to 3 November, the four-to-eleven-week extension, the redaction rights and guidance, the majority of the most informative evidence redacted, Anthropic's approval of every piece of information released, the cancelled compute allocation verification, the payment terms and the statement on misleading. metr.org
2. [2]Anthropic, PBC, Form S-1 Registration Statement, filed with the U.S. Securities and Exchange Commission on 26 October 2026. Source for the filing date, the quote on each Claude model accelerating the next, and the risk factor on lagging indicators of acceleration. sec.gov
3. [3]Anthropic, 1 June 2026. Anthropic confidentially submits draft S-1 to the SEC. Source for the confidential filing date. anthropic.com
4. [4]Anthropic, 15 October 2026. An update on independent evaluators. Source for the employee-like access given to METR and the statement that METR will publish its results. anthropic.com
5. [5]Yahoo Finance, 5 November 2026. Everything you need to know ahead of Anthropic's IPO. Source for the roadshow starting no earlier than 10 November under the SEC's 15-day rule, and the Wall Street Journal and Reuters reports of a debut as soon as the week of 16 November, as Yahoo Finance reports them. finance.yahoo.com