Anthropic is improving its AI models about twice as fast as it did a year ago, according to an independent review published Thursday. METR, the nonprofit group that evaluates frontier AI systems, spent 11 weeks inside Anthropic’s data and concluded the company is likely advancing on well-specified tasks at around double the pace it managed between March 2024 and October 2025. Its central estimate puts the speedup at 1.9 times the old rate. The researchers said they would “not be surprised” if the real figure landed anywhere from 1.3 to 3 times.
The bigger claim is what’s driving the speedup. METR’s subjective estimate is that most of it, perhaps around 80%, comes from Anthropic using AI to do AI research. That’s the dynamic Anthropic’s own chief executive, Dario Amodei, has spent months warning about in public, and it’s the first time an outside group with access to internal data has put a number on it.
“Many of their conclusions match my own. AI is speeding up our research, but it’s hard to say exactly how much,” Amodei said in a statement to the Associated Press.
The evidence behind those numbers is mostly hidden.
How METR measured Anthropic’s speedup
METR’s main yardstick is AECI, Anthropic’s internal version of the Epoch Capabilities Index, which folds benchmark scores into a single number. On that measure, Anthropic’s capability scores rose from about 15 points a year before October 2025 to 29 points a year after, roughly two years of historical progress packed into one. The 1.9x figure carries a 95% confidence interval of 1.5 to 2.8 times.
A second check came from Anthropic’s own published data. A trend-break fit in the Haiku 5.5 system card put the speedup at 1.53 times, with a 95% range of 1.20 to 1.82. METR said its higher figure reflects newer, unreleased models that the public card doesn’t cover.
From August 12 to October 29, about two and a half full-time METR staff worked on Anthropic devices with access to pretraining, reinforcement learning, compute, internal benchmarks, and speedup analyses, and interviewed roughly 15 employees. METR said it took no payment from Anthropic, which provided laptops and free tokens for the work.
METR said it is “confident” the jump wasn’t caused by faster growth in training compute. It also found no convincing evidence of bottlenecks in algorithmic progress, training compute, experiment compute, or data that would bite within the next 12 months. At the same time, it cautioned that AI is not yet capable of making highly significant, fully autonomous contributions to the most important stages of model training.
What Anthropic redacted, and why it matters
Anthropic had unilateral redaction rights over the roughly 100-page full report, and it used them. METR says the company “redacted a majority of the evidence we found most informative”: the per-stage acceleration estimates, the attribution breakdowns, the split between compute, data, and algorithms, bottleneck analyses, forecasts for future models, and the specific benchmark names.
Anthropic described the redactions as aimed at removing confidential information that could accelerate other AI developers. That’s a real concern for a company racing OpenAI and Google. But it leaves readers in an odd spot: the most striking claim in AI safety this year rests on evidence almost nobody outside Anthropic has seen.
METR published two documents on November 5: the independent assessment of AI research acceleration at Anthropic, and a separate review of the “Risks from automated R&D” section of Anthropic’s August 2026 Risk Report. The second document carries the sharper warning.
The threshold question
Anthropic’s Responsible Scaling Policy sets two thresholds for automated AI research. Its August Risk Report concluded the risk was low and that neither criterion was met as of July 15. METR agrees the risk from models as of that date was low. But it says Anthropic’s claim not to have crossed the 2x Acceleration Threshold, an early warning rather than a full stop, rests on “contestable judgment calls,” that it is “plausible” the threshold has already been crossed, and that the public report “overall understates the level of acceleration that was occurring.”
Anthropic told investors something similar in its own paperwork. An amended S-1 filing ahead of its planned Wall Street debut warned that its models may “imminently” cross the threshold for AI dramatically speeding up its own research. METR’s read of the data suggests the moment may already be here, or at least close enough that the difference between “imminent” and “already happened” is a judgment call, not a fact.
Amodei has publicly called for slowing the pace of AI development even as his company speeds up. The tension is hard to miss: the CEO asking the industry to hit the brakes is running a lab whose own evaluator says it’s now covering two years of ground in one.
Why this matters
If METR’s estimate is right, the safety clock is running faster than Anthropic’s public documents suggest. A lab that improves its models twice as fast as it did last year doesn’t just ship better products. It compresses the time between each safety review, each policy update, and each chance to catch a problem before the next model arrives.
Investors should pay attention too. Anthropic is heading for a public listing with a valuation debate already shaped by a brutal week for AI stocks. METR’s answer to the question of whether the AI boom can keep its pace is that the pace is accelerating inside Anthropic, at least, even as the market questions whether it can last.
None of this is settled fact. METR calls its headline number a high-uncertainty judgment, and it’s careful to say the acceleration applies to well-specified tasks, not everything AI does. The link between a benchmark index and real-world impact is uncertain, and the redactions mean no outsider can check the math.
Still, the direction is hard to argue with. An independent evaluator with 11 weeks of inside access says Anthropic’s AI is now improving about twice as fast, driven mostly by AI itself, with the evidence largely blacked out by the company being measured. For a field that keeps asking how fast is too fast, that’s the number everyone will be arguing about next.
Frequently asked questions
What did METR find about Anthropic’s AI progress?
METR estimates Anthropic is likely improving its models on well-specified tasks at about twice the rate it managed between March 2024 and October 2025. The central estimate is 1.9 times the old rate, with a 95% confidence interval of 1.5 to 2.8 times.
Is AI itself driving Anthropic’s speedup?
METR’s subjective estimate is that most of the acceleration, perhaps around 80%, comes from Anthropic using AI to do AI research. The group stresses that figure carries substantial uncertainty.
Has Anthropic crossed its recursive self-improvement threshold?
METR says it is plausible the company’s 2x Acceleration Threshold, an early warning under its Responsible Scaling Policy, has already been crossed. Anthropic’s August 2026 Risk Report said neither of its criteria was met as of July 15.
Why was so much of METR’s evidence redacted?
Anthropic had unilateral redaction rights over non-public information in the roughly 100-page report and removed a majority of the evidence METR found most informative. The company said the redactions were aimed at removing confidential details that could accelerate rival AI developers.
Sources: Reuters, Associated Press, METR, The Neuron, TechBooky