As of October 2026, Anthropic was likely improving its models on well-defined tasks at roughly twice its earlier pace, according to an independent METR assessment published November 5. The evaluators think AI-assisted research explains most of the current acceleration. They also attach substantial uncertainty to that estimate.
The comparison is with progress between March 2024 and October 2025. METR says a rate nearer 1.3 or three times that baseline would not surprise it. This measures how quickly capabilities improve, rather than saying a particular Claude model is twice as capable at every job.
METR examined internal development data and interviewed staff between August and October. It was told that the speedup mainly came from AI helping researchers, rather than fully autonomous contributions. Anthropic allowed extensive access, but redacted much of the evidence METR considered most informative from the public report, citing the risk of helping competitors accelerate.
That leaves readers relying heavily on the evaluators’ judgment. METR also cautions that results on verifiable tasks may not capture progress on broader, less clearly specified work.
Faster development changes the safety question
A separate METR review examined Anthropic’s August risk report, covering models deployed internally by July 15. It agreed that those models posed low direct catastrophic risk from automated research and development. Its disagreement concerned how confidently the public report dismissed an early-warning threshold for acceleration.
METR says crossing that threshold was plausible, although the relevant researchers at both organizations leaned toward it not having been crossed in July. The review argues that the public account understated the evidence for acceleration and the uncertainty in the private assessment.
The distinction matters: a warning threshold is meant to prompt scrutiny before a feared outcome occurs. It should not be reported as proof that the outcome has already happened.
Human supervision still matters
Anthropic’s own measurement framework offers a different view of the work. In its August snapshot, Claude led 26% of measured AI research and development, meaning it could perform most of a task from a high-level prompt while a human supervised. None of the measured subsets operated fully autonomously.
The company also proposes tracking how much agent activity gets monitored, how quickly reviews happen and how often actions are blocked or flagged. It acknowledges that using its own models as evaluators can leave shared blind spots, and points to independent verification as one remedy.
These reports concern different measurements and dates. Read together, they make the next useful question concrete: as AI takes on more research work, can outside evaluators check both the resulting progress and the oversight keeping pace with it?