The speed at which Anthropic’s AI models are improving has approximately doubled — and the evaluators who measured it think Anthropic’s own AI is probably the main reason why.

That is the central finding of a November 5, 2026 assessment by METR, an independent AI-evaluation nonprofit that spent roughly two and a half months examining internal Anthropic data. The report is careful about what it establishes: it is a measurement of benchmark progress rates at one company, made with significant uncertainty and published with key evidence redacted. It does not claim that AI is building itself, that catastrophe is imminent, or that every AI lab is moving faster. What it does claim — with meaningful supporting evidence — is that a feedback loop between AI tools and AI development at Anthropic appears to be real and strengthening.

What METR measured, and what it found

Between August 12 and October 29, 2026, a METR team of roughly 2.5 full-time equivalents conducted around 15 staff interviews, analyzed internal benchmark data, reviewed code and pull-request records, examined AI-agent usage logs, and received information about parts of Anthropic’s training pipeline. The assessment ran through October and the public report was released November 5.

The most concrete public number from that work is a 1.9× increase in the rate of growth of Anthropic’s internal model benchmark scores since October 2025, with a reported 95% confidence interval of 1.5× to 2.8×. METR translates this as roughly two years of historical benchmark progress compressed into one year. Its broader, higher-uncertainty headline judgment is that Anthropic was likely improving capabilities on well-specified tasks at about twice its March 2024–October 2025 pace by October 2026 — and that it would not be surprised if the true figure were anywhere from 1.3× to 3×.

It is worth being precise about what that means. The 1.9× figure is a comparison of growth rates, not a claim that model capability scores doubled, that the models are twice as useful, or that all AI development everywhere is moving faster. It describes how quickly Anthropic’s models are improving on a particular composite benchmark measure, compared with an earlier period at the same company.

To construct that measure, METR used an Anthropic-specific version of Epoch AI’s Capabilities Index, called AECI, which combines scores across multiple benchmarks into a single latent capability estimate using an item-response-style statistical model. The approach accounts for the fact that different benchmarks vary in difficulty and in how sharply they distinguish between models at different ability levels. METR fitted a piecewise trend with one breakpoint — estimated at October 7, 2025 — and reported the ratio of growth rates before and after it.

That method is sensitive to modeling choices. Alternative assumptions about how benchmark scores relate to underlying capability produced estimates ranging from 1.1× to 2.9×. When METR included models from other frontier labs in the analysis, the estimate fell to roughly 1.4×, though METR says there are competing interpretations of that result. Most other robustness checks — varying which model snapshots or benchmarks were included, or excluding benchmarks after they reached 85% accuracy — clustered around 1.9×. The reported confidence interval also reflects only the uncertainty from METR’s bootstrap procedure, not all uncertainty about which models belong in the trend fit. METR’s technical appendix details these caveats.

Why AI is probably the cause — and why that is hard to prove

Finding a faster trend is the easier half of METR’s claim. Attributing it to AI assistance is harder, and METR is explicit that this part of its conclusion is a judgment call, not a directly measured causal finding.

Its best guess is that AI use in R&D accounts for “perhaps ≈80%” of the current acceleration. METR stresses that this is an uncertain estimate informed by multiple lines of evidence, not a controlled experiment. Anthropic staff did not agree on a single explanation, and METR says non-AI factors — including hiring — probably played a significant role in the earlier stages of the acceleration, particularly in late 2025. The report does conclude that a faster rate of training-compute growth did not cause the acceleration, though the supporting details for that conclusion were redacted.

The public evidence METR can point to includes informal internal Slack polls in which Anthropic staff estimated how much AI sped up their work. For the most capable internal model available in October 2026, the average self-reported productivity uplift reached about 12×. METR is blunt about the limits of that number: polls may reflect an unrepresentative sample of staff and low-effort responses, and a person reporting they work faster does not mean the company’s overall capability progress accelerates by the same multiple. Translating individual productivity into model progress depends on how that work affects the training pipeline and what other bottlenecks exist.

METR also summarized interviews with 11 staff members who were asked whether AI had produced a greater-than-2× acceleration in their area of work. Two answered yes; three leaned yes; four gave borderline or unclear responses; two said no. Some interviewees described experiments becoming much easier to implement; others emphasized hiring or non-AI factors. METR used Claude to help classify interview quotes and notes that interviewees sometimes meant different things by “speedup,” conflating individual productivity with organizational acceleration. These accounts are evidence about staff experience, not a randomized causal study. METR’s discussion of this evidence includes the relevant qualifications.

Separate company-reported figures add context. Anthropic has said that in May 2026, more than 80% of lines merged into its codebase were authored by Claude — while acknowledging that lines of code are an imperfect productivity measure. In a separate description of its prototype R&D Automation Index, Anthropic reported that as of August 2026, Claude “leads” 26% of its AI R&D work, while the share at “AI collaborates” or above exceeded 90%; it reported no fully autonomous operation in the work it measured. These are Anthropic’s own measurements and definitions, not independent confirmation of METR’s causal attribution.

What “AI helping build AI” actually means here

The feedback loop METR describes works roughly like this: AI tools help Anthropic researchers write code, run experiments, and implement ideas more quickly. That assistance may help produce better next-generation models. Those better models, deployed internally as research tools, then help researchers even more — and so on.

That is meaningfully different from the stronger claim that AI is independently building its successor. METR is direct on this point: staff described the main current effect as AI assistance, not autonomous research. The report says AI is not yet making highly significant, fully autonomous contributions to the most important parts of model training, though researchers described early “signs of life” suggesting such contributions could grow.

An August 2026 preprint from researchers at Princeton, the UK AI Security Institute, Toronto, Berkeley, and other institutions offers relevant outside evidence. They gave AI agents six days, substantial compute, and two open-ended research questions drawn from unpublished NeurIPS submissions — with the task of producing work worthy of a top AI conference. The agents completed substantial engineering steps without human help but did not make meaningful progress on the central research questions. The original paper authors rejected both outputs. The researchers identified recurring problems with research judgment, responding to negative results, backtracking, and resource use. That study did not test Anthropic’s internal process, and its sample was small. But it illustrates why increased AI assistance on specified tasks should not be equated with autonomous open-ended research.

A related September 2026 working paper on the economics of recursive self-improvement models the conditions under which AI-driven R&D could become self-sustaining. Its rough calibration suggests a one-unit capability improvement would need to produce at least a 15% increase in AI R&D productivity to achieve self-sustaining acceleration; the authors’ back-of-the-envelope estimate of the actual return since coding agents emerged is around 9%. They conclude this does not yet establish self-sustaining acceleration, while noting that the return appears to be increasing. This is a model-based estimate, not a direct test of METR’s internal findings; two METR researchers are among its coauthors.

What the public cannot check

The assessment is high-access by industry standards — METR received internal benchmark results, training and reinforcement-learning information, algorithmic-efficiency measures, AI-usage data, staffing figures, and around 15 staff interviews. But it was not the more embedded, employee-like evaluator arrangement Anthropic announced separately in October 2026. METR did not receive query access to unreleased models and did not have unmediated access to internal documents and systems. Anthropic held unilateral redaction rights over the public report.

METR says Anthropic redacted a majority of the evidence it considered most informative for its risk conclusions. Its full report of roughly 100 pages was delivered to Anthropic but not published. METR explicitly warns that this leaves readers relying more heavily on its judgment than it considers ideal — meaning the public is, in practice, trusting METR’s summary of evidence it cannot inspect.

Anthropologic said redactions were intended to prevent disclosing information that could help competitors accelerate their own work. Anthropic co-founder and responsible-scaling officer Sam McCandlish was quoted by the Associated Press on November 5 acknowledging that the redactions made it harder for others to verify METR’s work: “AI is speeding up our research, but it’s hard to say exactly how much.” He said Anthropic planned to publish its own analysis of its pace of progress and RSP position in the coming weeks. As of November 6, that remained a stated future plan.

METR also discloses that it received no fee from Anthropic but used Anthropic devices and free tokens; one participating METR member had discussed and declined a job offer at Anthropic and held equity in a competitor; and Claude was involved in much of the data analysis. METR says an unconflicted team member thoroughly checked the potentially conflicted contributor’s work, and that it found no evidence of systematic distortion from Claude’s involvement. These disclosures do not demonstrate bias, but they are relevant to evaluating independence.

The companion dispute: was a safety threshold already crossed?

Alongside the acceleration assessment, METR published a companion review of Anthropic’s August 2026 risk report, which assessed models and activity through July 15, 2026 and described the risk from automated R&D as low.

METR broadly agrees that the models available by July 15 posed low direct catastrophic risk. But it says Anthropic’s public report understated the evidence for acceleration and that it was plausible the company had already crossed what its Responsible Scaling Policy calls the “2× Acceleration Threshold” — an early warning trigger, not a verdict that catastrophe has occurred or is imminent. METR’s own estimates for the relevant leading indicators ranged from 1.3× to 2.8×, depending on definition and period. Most researchers at both organizations leaned toward the threshold not having been crossed by July 15, but METR says the public report’s conclusion was overconfident given the evidence available at the time.

That distinction matters: an early warning threshold being plausibly crossed is not the same as finding that dangerous AI exists or that harm is coming. It is a prompt to look more carefully — which is precisely what the November assessment attempted to do.

METR did not find convincing evidence of near-term bottlenecks in algorithmic progress, training compute, experiment compute, or data that would clearly prevent further rapid progress within 12 months. It is explicit, however, that it is not confident in that assessment, that its understanding of bottlenecks changed as it received more information, and that some bottleneck analyses were redacted. The absence of an identified bottleneck is not evidence that none exists.

The policy context

In a September 2026 essay, Anthropic CEO Dario Amodei argued that the industry should “slow the pace at which we improve the capabilities of AI models,” proposing embedded evaluators, coordination among democratic-country labs, and possible global agreements. His argument is precisely that AI-assisted R&D acceleration may shrink the time available for safety work to keep pace. He also acknowledged the risks of unilateral slowdown and the difficulty of verifying international agreements — tensions METR’s findings do not resolve.

Anthropologic’s voluntary RSP and its October announcement of expanded evaluator access are governance steps, but neither the RSP nor the evaluator partnership is independently enforceable or automatically binding on other developers. METR’s acceleration assessment does not prove those safeguards are sufficient; it also does not prove they have failed.

What METR’s work does establish — within its stated uncertainty — is that the feedback loop between AI tools and AI development at Anthropic appears to be operating and may be strengthening. Whether that is fast enough to outpace safety work, and what should be done about it, are questions the evidence informs but does not settle.

Sources and method

All figures and characterizations in this article were checked against the METR primary sources. Ranges, confidence intervals, and hedges were preserved as reported. Claims attributed to Anthropic company sources are identified as company-authored. Forecasts, model-based estimates, and observed measurements are distinguished throughout.