METR Anthropic AI acceleration: about 2x, mostly redacted

On November 5, the research group METR published a public summary of a roughly 100-page report on how fast Anthropic is improving its own models, and the headline sentence is blunt. METR says Anthropic is likely improving capabilities on well-specified tasks at around twice the rate it managed between March 2024 and October 2025.

Anthropic gave METR unusual access to internal data. It also kept the right to decide what the public gets to see.

What did METR actually find?

METR staff, about 2.5 full-time people, spent eleven weeks at Anthropic between August 12 and October 29, 2026, with access to internal metrics, staff surveys and roughly 15 interviews. Their summary has three parts.

First, the speed-up. On Anthropic’s internal version of the Epoch Capabilities Index, called AECI, scores have grown 1.9 times faster since October 2025 than before it, with a 95% confidence interval of 1.5 to 2.8. In METR’s words, Anthropic made as much benchmark progress in the past year as it did in the two years before.

Second, the cause. METR guesses that perhaps 80% of today’s acceleration comes from Anthropic using AI to do AI research, though it calls that judgment highly uncertain. It says it is confident the speed-up did not come from faster growth in training compute.

Third, the brakes. METR found no convincing bottleneck in algorithms, compute or data that would bite in the next 12 months. It also says its view of those bottlenecks changed several times during the work, so it isn’t confident.

How sure is the 2x?

Not very, and METR says so repeatedly. Its own range is 1.3x to 3x, and different choices about which models and benchmarks to count gave 1.1x to 2.9x. METR also warns the AECI number “could be explained by overfitting to the benchmark distribution or otherwise be an artifact” of how the index is built.

The staff evidence is softer still. Anthropic’s Slack polls, where employees vote on how much faster the latest model makes them, reached 12x in October 2026. METR calls that “very loose evidence” and notes that a personal speed-up isn’t the same thing as the whole research pipeline speeding up. Interviews centered on about 2x. METR had Claude sort the interview quotes by whether they supported a 2x-or-higher speed-up: 2 yes, 3 leans yes, 4 borderline, 0 leans no and 2 no. METR reads that as mildly in favor.

What was blacked out?

Most of the evidence METR found most useful, by its own account. METR says Anthropic redacted “a majority of the evidence we found most informative” and quotes the company as saying the cuts remove confidential information that “could accelerate other AI developers.” Among the removed items: METR’s detailed estimates of how much AI speeds up each stage of model training, its analysis of specific bottlenecks, and the evidence that the speed-up wasn’t a compute story.

METR is direct about what that means. It writes that the public “has to rely on METR’s judgment much more than would be ideal.” Anthropic could reject any data request and held redaction rights over everything non-public, but METR kept the right to say that redactions happened and whether its findings depended on them. They did.

My read: that last point is the useful one. A reader can’t verify the 2x. What a reader can verify is that an outside group with real access believes it, and that the company chose not to show the evidence.

Where METR disagrees with Anthropic

The same day, METR published a separate review of the “automated R&D” section of Anthropic’s August 2026 Risk Report. Anthropic’s responsible scaling policy has an early-warning threshold at a doubling of the pace of AI progress that is substantially due to automating research. The Risk Report argued the line hadn’t been crossed by July 15.

Question	Anthropic’s Risk Report (August 2026)	METR (November 5)
Was the 2x early-warning threshold crossed by July 15?	No, acceleration was under 2x	Plausible that it was; the report’s claim looks overstated
Could models fully replace its research staff by then?	No	Agrees, calls the argument broadly compelling
Was there acute catastrophic risk from acceleration at that point?	Low	Agrees, calls it low

METR adds that most of its researchers and Anthropic’s lean toward “not crossed,” and that its own range of estimates, 1.3 to 2.8, was itself redacted from the public review. So this is a disagreement about judgment calls, not about a fact either side has shown.

What I make of it

My read: the number matters less than the direction of the argument. Anthropic’s own filings have already warned about recursive self-improvement, and I covered that in Anthropic’s IPO filing says it missed an AI capability jump before. METR’s report is an outside check saying the loop may already be turning, with the cause described as people using AI tools rather than AI running research alone. METR says that most of the work is still AI assisting staff, though researchers described “signs of life” for more autonomous contributions.

My read: I’d treat 2x as a decent guess from careful people, not as a measurement. And I’d treat the redactions as the story as much as the number, because a pace this fast is exactly where outsiders would want to see the data. What to watch is whether Anthropic publishes more of the unredacted material, and what its next Risk Report says about the threshold.

METR says it took no payment from Anthropic, though Anthropic provided free tokens for the analysis and METR worked on Anthropic devices. AECI is Anthropic's own index, so the benchmark side of this rests on the company's data.

Sources

Note on sources: I read both METR pages in full. Every figure, quote and the redaction details come from METR, which is the primary source here. I could not open Anthropic’s Risk Report PDF, so its claims are as METR quotes them. The Verge’s report adds Anthropic cofounder Sam McCandlish’s on-record response to the redactions and reaction from outside researchers; it doesn’t add new figures beyond METR’s own.