Anthropic is improving its AI about twice as fast as a year ago, outside group says

Anthropic is improving its AI models at roughly twice the pace of a year ago, according to an independent assessment published by the nonprofit METR on Thursday. METR's researchers say most of the speed-up likely comes from Anthropic using AI to help build its next models.

The timing is awkward. CEO Dario Amodei urged the industry to slow down in a Sept. 12 essay, and AP reporting carried by Scripps News says the company plans to go public later this month in what could be the biggest IPO ever.

What METR found

METR staff reviewed Anthropic's internal data from August 12 to October 29. They conclude that Anthropic is likely improving its models on well-specified tasks at around twice the rate it managed between March 2024 and October 2025.

The group puts that estimate under high uncertainty. It says it wouldn't be surprised if the true figure were closer to 1.3x or 3x.

The headline number comes from benchmark scores. Going by METR's report, Anthropic's models have improved at an estimated 1.9x their earlier rate since October 2025 (95% confidence interval: 1.5x to 2.8x). Put another way, the company made about two years of historical progress in the last year.

METR also says it is confident the speed-up didn't come from faster growth in training compute. It guesses that a majority, perhaps around 80%, of the current acceleration is due to AI, though it calls that judgment uncertain.

No obvious brakes

The part that worries METR most is compounding. If AI keeps speeding up AI research, each new model could make the next one faster still.

The researchers say they didn't find clear bottlenecks in algorithms, training compute, experiment compute or data over the next 12 months. They add that their view of those bottlenecks changed several times during the work, so they aren't confident in it.

For now, humans are still in the loop. The report says AI can't yet make highly significant, fully autonomous contributions to the most important parts of model training, but Anthropic researchers described "signs of life" suggesting that could change soon.

Redactions and caveats

Anthropic redacted a majority of the evidence METR found most informative, METR says. It described the cuts as aimed at removing confidential information that could speed up other developers.

Sam McCandlish, an Anthropic co-founder and its chief architect, responded on X. He acknowledged the redactions make it harder to "check METR's work", but said sharing the details risks handing out a roadmap for others to go faster. He also said Anthropic will publish its own analysis of its pace of progress in the coming weeks.

METR lists several reasons the number could be off. The gain may be limited to tasks that are easy to grade automatically, rather than "messy" real-world work. If METR plots frontier models from all developers instead of only Anthropic's, it gets a lower acceleration of 1.4x, because Anthropic's earlier growth looks unusually slow by comparison.

METR's report also says Anthropic's own Haiku 5.5 system card put the figure at 1.7x. METR says the gap to its 1.9x comes mainly from including newer, unreleased internal models.

One more disagreement: AP reports METR concluded that Anthropic's public August risk report "overall understates the level of acceleration that was occurring." METR also notes that its team had internal disagreements over the estimate.

The stakes for investors and policymakers

The findings matter because, as AP puts it, most information about AI risks has come from the companies building the systems, and the Trump administration has encouraged them to police themselves on safety. A third party with access to internal data is rare.

It's also a little awkward for Amodei's argument. In the September essay he described recursive self-improvement, AI's growing ability to build the next generation of AI, as something that "must be pursued very carefully, if at all." METR's report suggests Anthropic is already seeing a version of it.

Honestly, I'd treat the 2x figure as a rough signal, not a settled measurement. METR itself says the public has to rely on its judgment more than it would like, given how much was redacted.

I think the bigger story is the redaction. Prospective IPO investors will probably want to know what's behind the speed-up, and the people best placed to say aren't allowed to.

What happens next

Anthropic's own report on the pace of its progress is due in the coming weeks. The IPO is expected later this month. Anthropic didn't immediately respond to AP's request for comment.

METR says it has also published a separate review of the "Risks from automated R&D" section of Anthropic's August risk report, available on its site.