An outside auditor says Anthropic is now improving its models twice as fast, mostly because Claude is doing the research. Anthropic redacted most of the evidence.

METR, the Berkeley nonprofit that tests frontier AI models, published an independent assessment of AI research acceleration at Anthropic on 5 November. After eleven weeks with access to Anthropic’s internal data, it concluded that Anthropic is improving its models on well-specified tasks at about twice the rate it managed between March 2024 and October 2025. METR’s best guess is that around 80% of that speedup comes from AI doing AI research. In a separate review, it said Anthropic’s own public risk report from August understated how fast things are moving.

This is the scenario AI safety researchers have worried about for years: a model that helps build its successor, which then helps build the next one faster. METR does not say Anthropic has reached that point. It says the early version of it is probably already showing up in the numbers.

What METR measured inside Anthropic

METR’s central finding is that Anthropic’s internal capability index has grown about 1.9 times faster since October 2025 than before, from roughly 15 points a year to roughly 29. That is about two years of the old progress packed into one. The 95% confidence range runs from 1.5x to 2.8x, and different ways of fitting the data give anything from 1.1x to 2.9x.

The team of about 2.5 full-time staff worked from 12 August to 29 October. Anthropic gave it internal benchmark scores, training metrics, compute figures, headcount, inference spending and staff surveys, and METR interviewed about 15 employees. It did not get to query Anthropic’s unreleased models.

Three findings make the result harder to dismiss. METR says it is confident the speedup did not come from a jump in training compute. It found no clear bottleneck, in compute, data or algorithms, that would slow things down over the next twelve months, though it admits its view on that changed several times. And it estimates that AI accounts for most of the recent acceleration, where non-AI factors such as hiring explained more of the jump in late 2025.

Some of the evidence is soft, and METR says so. Informal Slack polls put the productivity boost from Anthropic’s best internal model at about 12x, which METR calls “very loose evidence”. Researchers told METR that AI is not yet making fully autonomous, highly significant contributions to the most important training work, but described “signs of life”.

Why METR says Anthropic’s public report understated the speedup

Anthropic’s Responsible Scaling Policy treats a doubling of AI progress attributable to AI research as an early warning. In its August risk report, Anthropic told the public the acceleration was “less than a factor of 2”. METR’s review says the unredacted private version implied 1.97, and that the 2x threshold was plausibly crossed by 15 July. It called the public report overconfident and asked Anthropic to change the wording.

The S-1 filing has moved in the same direction. In the amended prospectus filed on 2 November, Anthropic told investors that future models “may imminently meet” its threshold for dramatically accelerated progress, The Gadgeteer reports, while saying no current model has been designated as crossing it. The October version said Anthropic might not spot such an acceleration until after it happened. The new one says it has in the past not spotted one. We covered what the first filing disclosed in Anthropic’s admission of a training pause and escaped models.

Sam McCandlish, Anthropic’s co-founder and responsible scaling officer, replied on X that many of METR’s conclusions match his own. He wrote that “AI is speeding up our research, but it’s hard to say exactly how much”, according to the Associated Press.

How much of the evidence the public can check

Very little. Anthropic had the right to redact the report, and METR says it removed most of the evidence METR found most informative. METR’s own summary is blunt: “the public has to rely on METR’s judgment much more than would be ideal.” Anthropic says the cuts were meant to stop the data from helping competitors move faster, and McCandlish defended them as commercially sensitive while agreeing they make the work harder to verify.

Anthropic also cancelled one part of the project halfway through: checking how it splits compute between capability work and safety research. The Decoder reports that Anthropic said it wanted to focus on risk assessment first. Former OpenAI researcher Daniel Kokotajlo criticised the decision, because that audit is what would make a promise to slow down verifiable.

That matters for the debate Dario Amodei himself started. In September he asked the industry to slow down so safety work could keep up. A lab that asks for pacing and then removes the one audit that would show whether it is pacing itself is asking to be trusted on the most important number.

The case that METR is overstating it

There is a serious skeptical reading, and much of it comes from METR. The acceleration applies to well-specified tasks with clear right answers, which AI is good at; the messy work of choosing which ideas to test may not speed up the same way. When METR adds frontier models from OpenAI, Google and xAI, the industry-wide speedup drops to about 1.4x, which could mean Anthropic was simply catching up from an unusually slow baseline.

Inside Anthropic, views differ. One researcher told METR that AI explains only 15% to 20% of the productivity gains and that colleagues’ self-reports should not be trusted; another named new hires as the main factor, The Decoder reports. Karthik Tadepalli, who argues on LessWrong that the acceleration is probably real, points out that Anthropic’s own Opus 5.5 system card found a one-time jump fitted the data better than a lasting change of trend. Nobody outside can settle that, because the data is redacted.

My read: the direction is not in doubt, and the size is. Even the low end of METR’s range means the model you sign a contract for today will be outdated sooner than you planned. And both METR and Anthropic rate the immediate risk of catastrophe from current models as low, while METR warns it could rise to high within six months if the acceleration continues.

What European companies buying AI should do now

Plan for shorter model lifetimes. If Anthropic’s progress has roughly doubled, the gap between a model you validated for a process and the version the vendor wants you on will shrink. Build your evaluation set now, the fifty or hundred real cases your process depends on, so that testing a new version takes days rather than a project.

Ask vendors for outside evaluations, and read what was cut. METR’s report is the most detailed independent look inside a frontier lab so far, and it still leaves the public relying on trust. When a supplier points to third-party audits in a security review, ask who did them, what access they had and what was redacted.

Watch the EU’s reaction. Under the AI Act, providers of general-purpose models with systemic risk must assess and mitigate those risks and report serious incidents to the European Commission’s AI Office, which has already opened its first case against OpenAI and Anthropic. A finding that a lab’s own warning threshold was plausibly crossed in July, without the public report saying so, is the kind of thing the AI Office can ask about. Anthropic’s listing on Nasdaq, expected this month, will add securities lawyers to the readers, and as the leaked prospectus showed, the risk section is where Anthropic says things it does not say elsewhere.