METR’s assessment of AI R&D acceleration at Anthropic

METR. Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, 2026-11-05.

Is AI already speeding up the research that builds the next AI? The answer bears on whether the step from AGI to Superintelligence takes years or months, and the evidence sits inside the labs. From August to October 2026, a small METR team worked with Anthropic’s internal data: benchmark scores for unreleased models, measures of algorithmic progress, compute figures, staff surveys and about fifteen interviews. It is the most detailed outside measurement so far of Recursive self-improvement at a frontier lab, and the answer was larger than many expected.

What it found

• Progress is about twice as fast. As of October 2026, Anthropic was likely improving its models on well-specified tasks at about twice the rate of March 2024 to October 2025. METR “would not be surprised” by anything from 1.3x to 3x. The most visible evidence is a benchmark index, Anthropic’s version of the Epoch Capabilities Index, which has grown 1.9 times faster since October 2025 (95% range 1.5–2.8): roughly two years of the old progress in one.
• Most of the speed-up is probably AI. METR guesses that about 80% of the current acceleration comes from Anthropic using AI in its own research. It is confident that faster growth in training compute is not the cause. Staff report large personal speed-ups, reaching 12x in informal polls, but say AI mostly assists them and cannot yet make the most important contributions on its own. They describe “signs of life” that it soon will.
• No bottleneck in sight. METR found no convincing limit from algorithms, compute, experiments or data over the next twelve months, so it sees a significant risk that progress speeds up further. Its view of the bottlenecks reversed several times during the work, so it holds this loosely.

A companion review of Anthropic’s August Risk Report adds that Anthropic’s own public account understated the acceleration. Anthropic’s safety policy treats a twofold speed-up from automated research as a warning threshold. Both METR and Anthropic’s staff lean towards it not having been crossed by mid-July, but METR thinks it plausibly had been.

How far to trust it

• Most of the evidence is hidden. METR wrote a report of about 100 pages; Anthropic redacted most of the evidence METR found most informative, citing the risk of helping other developers. Outsiders can check the benchmark trend, but the attribution to AI and the absence of bottlenecks rest largely on METR’s judgment.
• The tasks are the clean ones. The benchmarks measure well-specified, automatically graded work. Progress on messy real-world work may be slower.
• The baseline may have been low. Anthropic’s models improved more slowly than the industry frontier before late 2025, which could make the later speed-up look bigger than it is.
• Or it may be an underestimate. The growth rate itself may be rising, fuller automation may be close, and benchmarks have sometimes understated real-world impact.

Anthropic also cancelled a planned part of the review that would have checked how it divides compute between capabilities and safety, saying it would revisit it later.

Responses

• Anthropic agrees, and defends the redactions. Sam McCandlish, an Anthropic co-founder who leads much of its safety work, wrote that many of the conclusions match his own. Sharing what drives the speed-up “risks handing out a roadmap for others to go faster”, and is commercially sensitive. Anthropic plans its own report on its pace of progress.
• The redactions are self-serving. Daniel Kokotajlo replied that citing others’ speed is “pretty rich” from a lab that is not slowing down, and that Anthropic “appears to be recklessly gambling with our lives”. See Should safety-focused labs keep building.
• A big update towards short timelines. Ryan Greenblatt, an alignment researcher who also works at METR, called it “a big update towards ASI within 1-2 years”. He had expected a speed-up of only 1.15–1.3x in 2026. Because some progress comes from compute rather than better methods, and because ideas get harder to find, he estimates that AI is now speeding up the improvement of algorithms and data about 4.5 times, a level AI 2027 reaches around March 2027.
• A skeptic concedes. Jake Mendel, who had longer timelines than his friends at Anthropic, wrote that if METR is right, they were “much more right about timelines than me”. He now puts more than 50% on superintelligence within 18 months, absent a slowdown.

What it changes

The report does not settle how fast superintelligence could follow, but it moves the argument from mechanism to measurement. A feedback loop is not a speed estimate said that a feedback loop needs a speed estimate; this is the first public one. The step from AGI to superintelligence could be short named a speed-up before full automation as the clearest early signal; this is that signal. Both sides should now watch whether the rate keeps rising.

Related: AGI timelines, Intelligence explosion, AI takeoff.