What happened
On November 5, 2026, METR published an independent assessment of AI’s contribution to research and development at Anthropic. METR staff conducted the high-access evaluation from August 12 through October 29, 2026. Anthropic shared internal information spanning pretraining and reinforcement learning, training compute, internal-model benchmark scores, analyses of AI-related speedups, forecasts of a future model’s metrics, and interviews with relevant staff. METR describes this as unprecedented access to internal data about Anthropic’s AI R&D process.
METR’s headline judgment is that, as of October 2026, Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate it did between March 2024 and October 2025. The report stresses substantial uncertainty: METR says it would not be surprised if the true acceleration were closer to 1.3 times or three times the historical rate. It is also less certain how the result transfers from well-specified tasks to broader, less well-specified work.
One important empirical input is the growth of scores on AECI, Anthropic’s version of the Epoch Capabilities Index. Since October 2025, METR estimates that those scores grew at 1.9 times the pre-October 2025 rate, with a 95% confidence interval of 1.5–2.8 times. METR says this implies roughly two years of historical progress over the last year. It is confident that the acceleration was not caused by an acceleration in the scaling of training compute. However, the report notes that AECI may overstate or understate real-world progress: its results could reflect benchmark-distribution overfitting or limitations in the index’s methodology.
METR’s interviews and surveys are consistent with AI use contributing to a doubling in the pace of capability progress, but it cautions that self-reported uplift is difficult to interpret. It says R&D acceleration appears to come primarily from AI assistance, rather than fully autonomous contributions, while there are signs that fully autonomous contributions could play a larger role soon. METR also found no evidence of significant near-term bottlenecks from algorithmic progress, training compute, experiment compute, or data. It warns that its view of bottlenecks shifted as it received more information, so this conclusion is uncertain.
The assessment has an unusual transparency limit: Anthropic had the right to redact non-public information, and redacted a majority of the evidence METR found most informative for its conclusions. METR says some redacted material is hard to describe without disclosing sensitive information. It concludes that readers must rely on METR’s judgment more than would be ideal, and that the report does not achieve transparency about the underlying evidence. The public post provides no product launch, pricing change, or new access terms.
What the archive adds
This is a new finding on a related METR thread, not a restatement of prior coverage. METR explains its approach in a companion post described why process disclosure matters when interpreting its public AI-risk work. An earlier METR alignment assessment focused on how evaluation should probe possible agent misbehavior. The new report brings that concern into a specific, high-access assessment of AI’s contribution to one lab’s R&D pace.
Trend signal
No supplied SITE_PATTERNS thesis directly captures this finding. METR’s report concerns the pace and possible causes of capability progress inside a frontier lab; it does not establish a general operating model or a durable product boundary for agent systems.
Why now
METR’s assessment covers activity through October 29, 2026, and reports a change in estimated capability-progress rates since October 2025. Its timing makes the post a recent retrospective on the preceding year, rather than a projection based only on future scenarios.
Connections to other events
The new assessment connects to METR’s earlier public work on how to evaluate risks from advanced AI systems. Its companion post on assessment process makes transparency relevant to interpreting these results; its prior alignment discussion asked what evaluations should reveal about possible loss of human control. This report adds evidence about a different part of that picture: whether AI use inside R&D could speed the development of future capabilities.
Why it matters
For teams building with AI, the report is a reason to distinguish “AI helped us do more work” from “AI accelerated the system’s capability frontier.” METR’s evidence suggests that, in Anthropic’s case, assistance to researchers may already be contributing materially to faster progress, even though the report does not say that autonomous AI researchers are now doing most of the work. If the pattern is real, the pace of model change may depend not only on compute and algorithmic advances but also on how effectively labs use existing models in their own research loops.
That is a consequential signal, not a ready-made forecast for every company. METR’s estimate applies to Anthropic and well-specified tasks, and the public cannot inspect much of the evidence behind its judgment. Teams should therefore avoid treating “twice as fast” as a universal multiplier. The practical implication is to track AI’s contribution to research workflows separately from output volume, and to preserve enough measurement and review to tell genuine acceleration from benchmark effects or other inputs.
Limits and uncertainty
The report does not establish that every AI lab, model family, or real-world task is progressing at the same rate. METR gives a wide plausible range, flags uncertainty about how AECI maps to broader impact, and acknowledges possible non-AI explanations for some progress. Its lack of observed bottlenecks over the next 12 months is also explicitly uncertain. Most importantly, Anthropic redacted much of the evidence METR considered most informative, reducing the public’s ability to independently assess the analysis. The assessment is not proof that AI systems are autonomously conducting R&D or that acceleration will continue.
Skill to build now
For one AI-assisted research workflow this week, record a baseline before adding more automation: define the task, its quality check, and the time researchers spend on it. Then compare the same measures after introducing AI assistance, keeping human review and other workflow changes visible. This will not reproduce METR’s assessment, but it can help a team separate faster drafting or experimentation from verified gains in research throughput.
Reader takeaway
METR’s high-access assessment suggests Anthropic’s well-specified model capabilities have been advancing about twice as fast since October 2025, with AI assistance likely contributing. The result deserves attention, but the estimate is uncertain and much of its strongest evidence is redacted. Treat it as a serious signal about AI-enabled R&D, not as a universal growth rate or proof of autonomous research.