Jared Kaplan, Anthropic’s co-founder and chief science officer, in an official company portrait. Image: Anthropic
Anthropic is now improving its AI models about twice as fast as it did before late 2025, and most of that speed-up likely comes from using its own AI to do the research. That is the verdict of METR, the nonprofit AI evaluator, in an independent assessment published on Nov 5 after its staff spent 11 weeks examining Anthropic’s internal data. But Anthropic redacted most of the evidence METR found most convincing, so readers have to take the group’s word for it.
METR says Anthropic is “likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025.” It would not be surprised if the true figure were closer to 1.3 or 3 times the old rate.
Two years of progress in one #
From Aug 12 to Oct 29, METR staff worked on Anthropic devices with data on pretraining, reinforcement learning, compute, internal benchmarks and speed-up analyses, plus about 15 staff interviews. METR took no payment. The central figure comes from AECI, Anthropic’s version of the Epoch Capabilities Index, which combines benchmark scores. Since October 2025 it has grown at 1.9 times its earlier rate (95% confidence interval: 1.5 to 2.8 times), which METR says means Anthropic made “roughly two years of historical progress over the last year”, about the distance from Claude 2.1 to Sonnet 4.5. The Claude Haiku 5.5 system card said 1.7 times; METR also counted three unreleased models trained since July.
METR is “confident” the jump was not caused by faster growth in training compute. It guesses that “perhaps ≈80%” of the current acceleration is due to AI, though it thinks other factors played a large role in late 2025. In Slack polls, staff estimates of their own speed-up from the latest model reached 12 times in October, which METR calls “only very loose evidence”.
“Signs of life” #
AI cannot yet make “highly significant and fully autonomous contributions” to the most important parts of training, METR says, but researchers described “signs of life”. Anonymised interview notes range from sceptics who credit hiring to enthusiasts like this one:
I’ve seen us discovering very promising directions ~entirely because people can spam 10 ideas at [agent-scaffold] and see what sticks.
An Anthropic researcher, in interview notes published by METR
Another guessed only 15 to 20% of the change was due to AI and said they don’t “believe in self-reported productivity increase”. METR found no clear bottleneck in algorithms, compute or data likely to slow things in the next 12 months, so it sees “a significant risk that further capabilities progress could lead to even faster acceleration”.
Claude was “deeply involved” in most of the data METR received and in its own analysis, though METR found no sign that skewed the results.
Most of the evidence is redacted #
Anthropic had the right to redact any non-public information, and used it. METR says Anthropic “redacted a majority of the evidence we found most informative”, which Anthropic described as “aimed at removing confidential information that could accelerate other AI developers.” Algorithmic progress estimates, forecasts for future models and breakdowns by training stage were all cut. As a result, “the public has to rely on METR’s judgment much more than would be ideal, given the weight of the questions answered here.”
METR also says Anthropic changed the scope midway to drop a planned check on how much of its compute goes to capabilities versus safety, saying it would revisit that later. This was a separate engagement from the embedded evaluators Anthropic announced on Oct 15, and such checks remain voluntary, as Anthropic’s public benefit report admitted. Anthropic hasn’t commented on the findings.
An early warning that may already have gone off #
METR’s companion review of the automated R&D section of Anthropic’s August 2026 Risk Report agrees risk from models as of July 15 was low. But it says Anthropic’s claim not to have crossed what METR calls its “2x Acceleration Threshold”, an early warning, rests on “contestable judgment calls”, that it is “plausible” the threshold had been crossed, and that the public report “overall understates the level of acceleration that was occurring.”
Anthropic’s chief executive has warned about exactly this. In his September essay, Dario Amodei wrote: Since roughly this summer, AI has been advancing drastically faster, driven primarily by AI’s growing ability to build the next generation of AI.
Dario Amodei, Anthropic chief executive
Why it matters #
This is a rare outside, data-backed estimate that a frontier lab’s AI is speeding up its own development, the dynamic Amodei says “must be pursued very carefully, if at all.” It lands as Anthropic heads for a Nasdaq listing, and with the key evidence redacted, the public can check little of it.
Sources: METR, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026; METR, review of the automated R&D section of Anthropic’s August 2026 Risk Report; Anthropic, Risk Report, August 2026 (redacted); Dario Amodei, We Must Pace the Frontier.