METR finds Anthropic's AI research is moving about 1.9 times faster, and the best evidence is redacted

METR, the nonprofit that measures how long AI systems can work on their own, has published the first outside look at whether AI is speeding up AI research inside a frontier lab. Its answer for Anthropic, dated November 5, 2026, is yes, by roughly 1.9 times on the main metric. The catch is that Anthropic decided what the public gets to see.

What METR measured

The report is titled Independent Assessment of AI R&D Acceleration at Anthropic in October 2026. METR staff, about 2.5 full-time equivalents, ran a high-access evaluation between August 12 and October 29, 2026, and submitted the final report on November 3.

The main quantitative tool was a piecewise linear fit to AECI, Anthropic's internal version of the Epoch Capabilities Index, with a breakpoint at October 7, 2025. Since that date, METR finds growth about 1.9x the prior rate, with a 95% confidence interval of 1.5 to 2.8x. Its own robustness checks range from 1.1x to 2.9x.

METR also says Anthropic is likely improving its models on well-specified tasks at around twice its March 2024 to October 2025 rate. The report adds that "we would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate." That is an honest error bar, and it is wide.

How much is AI doing the speeding up

METR guesses that roughly 80% of the October 2026 acceleration is due to AI. It labels this a subjective estimate with substantial uncertainty. The other 20% would come from factors such as more compute, more staff and better data, which are also hard to separate.

Two softer lines of evidence point the same way. Slack polls of Anthropic staff reached about 12x self-reported uplift for the most capable internal model in October 2026, which METR calls "very loose evidence." And in about 15 staff interviews, 11 interviewees could be classified: two supported more than 2x acceleration, three leaned yes, four were borderline or unclear, none leaned no, and two said no.

One anonymized researcher told METR "It is hard to see how I am not being 3x more productive." Another guessed that only 15 to 20% of the change was due to AI and did not believe self-reported productivity gains. The disagreement inside the building is part of the finding.

The redaction problem

Anthropic had unilateral redaction rights over the report. METR says redacted items include detailed per-stage acceleration estimates, bottleneck analyses, forecasts for future models and the names of specific benchmarks. Anthropic described the redactions as "aimed at removing confidential information that could accelerate other AI developers."

METR also reports that a planned check on how Anthropic allocates compute was canceled midway through the engagement, and that "Much of the redacted evidence we viewed had multiple plausible interpretations." Anthropic co-founder Sam McCandlish said on X on November 5 that "We redacted a lot, and I accept that this makes it harder for others to check METR's work," according to FourWeekMBA's summary.

This matters because the claim is about whether a safety threshold has been crossed. Anthropic's amended IPO filing of November 2 said, as summarized by FourWeekMBA, that it "may imminently meet" the "dramatically accelerated progress" definition in its Responsible Scaling Policy without designating a current model as exceeding it. Readers cannot check that reading without the redacted material. Our earlier look at METR's review of the August risk report found the same pattern.

Where the method is weak

METR lists its own limits. AECI may overfit benchmarks and depends on the link function. The report says one internal model scores above 50% on all 15 benchmarks used, which raises concerns about ceilings, since a saturated test cannot show acceleration. Non-AI factors may explain much of the late 2025 jump, and Anthropic's history may have been unusually slow.

There are also process disclosures. Claude was heavily involved in data collection, writing and analysis. One METR team member had a prior equity stake in a competitor and had declined a job offer from Anthropic, so their work was independently checked. METR writes that "we think it is extremely unlikely Anthropic was actively trying to mislead us."

What it means for the race

Outside analysts are already reading the number hard. Ajeya Cotra wrote on November 6, in a post she says does not speak for METR, that "it looks like the 'dramatic acceleration' threshold has already been crossed," per FourWeekMBA. METR's own staff reportedly lean toward the opposite on one specific date, July 15, and The Hacker News reports METR's summary line that Anthropic is "likely improving the capabilities of their models on well-specified tasks at around twice the rate."

The honest reading is narrower than either camp wants. Benchmarks on well-specified tasks are growing faster inside one lab, and its own staff say AI tools are a large part of why. That is not the same as autonomous self-improvement, which our explainer on recursive self-improvement separates into strict and loose versions. This evidence supports the loose one.

Our take

This is the most useful public data point yet on AI accelerating AI work, and also a case study in why labs should not control the evidence about their own acceleration. We would ask for three things: Anthropic's own analysis, which McCandlish says is due in the coming weeks; a redaction log that says what category each cut falls in; and a repeat of the compute-allocation check that was dropped. Until then, 1.9x is a number to hold loosely.