What happens when AI tools help researchers develop their successors? A new METR assessment of Anthropic puts that feedback loop under scrutiny.
The important question reaches beyond any single model launch: if research tools improve with the models they help produce, how should companies measure the resulting feedback loop?
What METR actually measured
Published on November 5, METR’s assessment draws on access to Anthropic’s internal research data and staff between August 12 and October 29, 2026.
Its central judgment is that capability improvement on well-specified tasks is probably progressing at roughly twice the earlier rate, with substantial uncertainty. METR considers approximately 1.3 to 3 times the historical pace plausible.
The researchers believe AI use likely explains most of the acceleration. They describe the current contribution primarily as assistance to researchers, rather than fully autonomous research.
Why a faster research cycle matters
Consider an illustrative research team testing a training idea. An assistant could help write experiment code, inspect results and prepare follow-up tests. If those steps take less time, the team could investigate more ideas within the same period.
Some experiments would still fail. Others might improve the next model, which could then become a more useful research assistant. This is the feedback loop that makes AI-assisted AI development worth watching.
That mechanism alone does not establish unlimited acceleration. In this example, an unresolved scientific question, an unreliable evaluation or a delay in obtaining computing resources could still determine when the project finishes.
It also helps to separate three questions: how quickly researchers complete tasks, how quickly model capabilities improve, and how quickly customers receive useful products. A change in one does not automatically quantify the others.
Outside evaluation still leaves a visibility gap
METR says Anthropic redacted most of the evidence it found most informative. Readers can inspect the headline conclusions, but cannot independently examine much of their foundation.
In comments reported by the Associated Press, Anthropic co-founder Sam McCandlish broadly agreed with many conclusions while emphasizing uncertainty about the size of the speedup. He defended withholding details on commercial and safety grounds, while acknowledging that redactions make external checking harder.
The governance trade-off is clear: protecting sensitive methods can be reasonable, while leaving the public dependent on an evaluator’s judgment. Our reading is that an assessment’s independence and the public reproducibility of its findings should be considered separately.
A safety warning is a judgment about evidence
A separate METR review examines the automated research section of Anthropic’s August risk report.
METR agrees that models assessed as of July 15 posed low acute catastrophic risk from research acceleration. However, it argues that Anthropic’s public account understated the acceleration underway and did not clearly communicate the contestable judgments behind its conclusion about an early warning threshold.
METR considers it plausible that this threshold had been crossed, although both its researchers and the relevant Anthropic staff lean toward it having remained uncrossed.
For readers, this distinction matters. An early warning threshold is intended to prompt scrutiny before the most serious outcomes occur. Debate over that threshold should therefore be evaluated on the evidence and the decision process, without treating it as a declaration that catastrophe is imminent.
What this means for teams using AI
For a business considering a new assistant, our practical recommendation is to keep the scope of the evidence visible. Results inside a frontier AI laboratory should motivate questions about your workflow, rather than become a productivity forecast for your organization.
Ask where assistance would help, what counts as an acceptable result and who can verify it. A faster draft may be useful; the meaningful operational outcome is a finished task that meets the required standard.
DemystIA’s guide to interpreting AI benchmarks offers a way to connect evaluation results to actual work. Our article on the human attention bottleneck explores the review work that can accompany greater automation.
The development to watch is whether future assessments can connect better research tools to repeatable capability gains, while providing enough evidence for meaningful external challenge. That would make the debate about AI acceleration more measurable—and more useful.
Sources
- METR — Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, November 5, 2026. Primary assessment.
- METR — Review of the “Risks from automated R&D” section in Anthropic’s Risk Report (August 2026), November 5, 2026. Separate risk-report review.
- Associated Press — Anthropic is rapidly accelerating AI development amid slowdown calls, outside research group says, November 5, 2026. Reporting and Anthropic’s response.