On November 5, 2026, the AI evaluation organization METR published an independent assessment based on Anthropic's internal data and other sources. It concludes that the pace of capability improvement in Anthropic's AI models had accelerated to roughly double its earlier rate as of October 2026. On a capability index that combines multiple benchmarks, METR estimates the pace reached 1.9 times the historical rate. However, the fact that capability gains accelerated is a separate question from whether the cause was the automation of AI research and development. Nor can this figure alone tell us whether Anthropic crossed the "2x" early-warning threshold set out in its safety policy. Reading it together with METR's review of Anthropic's August 2026 risk report, published the same day, shows the need to distinguish three issues: the speed of capability gains, their cause, and whether a safety threshold has been reached.
Capability gains at 1.9x the earlier pace, distinct from the safety policy's "2x threshold"
The "AECI" that METR used in its analysis is an index that estimates AI model capability from multiple benchmark results. It is the Anthropic version of the "Epoch Capabilities Index." The 1.9x figure compares the rate at which AECI rises over time with the rate in the past. It does not mean that benchmark scores or AI intelligence itself became 1.9 times higher. METR explains that the pace is now such that benchmark capability gains that previously took about two years are achieved in about one year.
For the analysis, METR fit straight lines to the AECI trajectory before and after October 7, 2025, and estimated the improvement pace from the difference in slopes. The models evaluated include unreleased ones, and the date used for each model is not its public release date but the date it became available to most Anthropic employees. The result therefore does not measure how often new products reached general users; it assesses the progress of model development inside the company.
The 95% confidence interval for the main estimate of 1.9x was 1.5 to 2.8x. However, this range reflects only the uncertainty in each model's AECI estimate, not all methodological uncertainty, such as which models to include in the analysis. In fact, when the analytical method was changed, the estimates spread to between 1.1 and 2.9x. METR's overall judgment, using March 2024 to October 2025 as the baseline, also puts the acceleration at about 2x, though it acknowledges the true figure could be around 1.3 to 3x. These numbers rest on different methods and judgments; they are not separate expressions of error around a single estimate.
Meanwhile, Anthropic's safety policy, the "Responsible Scaling Policy (RSP)," sets out a separate criterion for detecting acceleration of capability gains caused by automation of R&D. This early-warning threshold evaluates whether the pace of capability improvement has doubled relative to both the growth rate expected if AI had not made a significant contribution to research and the fastest sustained growth rate observed in the past. The latter covers at least three model generations, and the judgment also requires that a substantial part of the acceleration be attributable to AI R&D automation and that further acceleration be expected.
The 1.9x that METR calculated from the capability index as of October 2026 and the question of whether Anthropic exceeded the safety policy threshold as of July 15, 2026 therefore differ in both comparison baseline and evaluation date. Organizing METR's assessment of capability gains and its review of the August risk report gives the following picture.
| Basis of comparison | Evaluation date | Main result or judgment |
|---|---|---|
| AECI improvement pace compared with the past | Through October 2026; analyzes changes since October 2025 | Main estimate 1.9x; 95% confidence interval 1.5–2.8x. This does not include the uncertainty of the analysis method as a whole |
| RSP determination based on the estimated growth rate without AI and the fastest sustained past growth rate | July 15, 2026 | The 2x threshold may have been exceeded, but many of the METR and Anthropic researchers involved lean toward the view that it had not been reached |
Thus, even though both evaluations deal with the same "2x" number, its meaning differs. An AECI improvement pace of 1.9x does not allow us to conclude that the RSP threshold was narrowly missed. Conversely, the fact that the confidence interval extends above 2x is not by itself grounds for judging that the safety policy threshold was crossed.
How much of the AI R&D acceleration could be confirmed?
The assessment was conducted over about 11 weeks, from August 12 to October 29, 2026. METR committed 2.5 full-time-equivalent staff and conducted about 15 interviews with key Anthropic personnel. It also examined internal information on pretraining and reinforcement learning, capability metrics for unreleased models, and employee surveys. METR obtained access to information that Anthropic does not normally make public, but the results are not a peer-reviewed paper; they are an assessment report published by METR itself.
From its analysis, METR expresses strong confidence that faster growth in the compute devoted to training did not cause this acceleration in capability gains. It also investigated non-AI factors such as increased hiring, improved data, and algorithmic discoveries. It believes, however, that non-AI factors likely contributed substantially to the acceleration observed in late 2025.
So how much did AI itself contribute to the acceleration? METR speculates that most of the acceleration observed as of October 2026, probably about 80%, may be due to the use of AI. But this is a subjective estimate based on multiple pieces of circumstantial evidence and carries large uncertainty. METR did not rigorously separate AI use from other factors through experiments, and it did not obtain decisive evidence of causation. Moreover, the 80% is an estimate of AI's contribution to the "acceleration" of capability gains, not a claim that 80% of overall AI model capability was produced by AI.
In interviews with researchers, estimates of how much faster AI made their work were widely distributed, centered on about 2x. By contrast, in a vote on Anthropic's internal Slack, a self-reported 12x speedup in individual work with AI was recorded for October 2026. METR, however, rates the reliability of this vote as evidence as limited, since it was a quick survey whose responses did not require careful consideration. An individual researcher finishing work faster is not the same as an organization's overall AI development capability improving at the same rate.
In METR's view, what mainly supports the current acceleration in R&D is AI assistance to human researchers. It judges that AI has not yet reached a stage where it makes extremely large contributions to major model training processes independently of humans. It adds, however, that signs of such autonomous contributions possibly expanding in the future have also been observed, and that maturity differs by research area.
Anthropic's August risk report assessed acceleration as "below 2x"
The other review METR published the same day examined the R&D automation section of the risk report Anthropic released in August 2026. The situation evaluated was as of July 15. METR broadly agreed with Anthropic's conclusion that AI models at the time were not at a stage to fully replace human researchers and that the risk of catastrophic outcomes arising directly from R&D automation was low. At the same time, it judged that the "2x" early-warning threshold in the safety policy may already have been exceeded. A low direct risk and possibly having reached a threshold meant to warn of future risk are not necessarily contradictory.
What METR took issue with was how uncertainty was conveyed in Anthropic's public report. The public version described the acceleration rate of a leading indicator of capability gains as "below 2x." But when METR calculated it by combining the current and past growth rates given in the private version, the acceleration rate came to 1.97x. METR points out that for this value, "about 2x" would be a more appropriate description than "below 2x." Note that the 1.97x figure was not stated directly in the private version; it is a value METR calculated itself.
Because the underlying numbers used in the calculation have not been made public, ordinary readers cannot reproduce the same calculation. Moreover, even in METR's own analysis, estimates of the acceleration rate range from 1.3 to 2.8x depending on the method. The issue is not simply how the number is rounded, but whether such uncertainty was adequately conveyed in the public report.
According to METR, the private version of the report contained a caveat that the R&D acceleration rate relative to a no-AI scenario could possibly exceed 2x. In the public version, however, the judgment attributing the acceleration to non-AI factors was expressed with greater confidence. METR also points out that the AECI analysis at the time could not adequately distinguish whether the capability index had risen sharply but temporarily, or whether the improvement pace had become faster on a sustained basis.
However, METR did not formally certify that the early-warning threshold was crossed. Many of the METR and Anthropic researchers involved in this review lean toward the view that the threshold had not been reached as of July 15, 2026. What the review questions is less the judgment that the threshold was not reached than the fact that the possibility of a different conclusion, and the uncertainty accompanying the judgment, were not adequately explained in the public version.
Most of the evidence in the independent assessment is not public, leaving challenges for third-party verification
By having an outside organization like METR examine Anthropic's internal information, the material available for judgment has certainly grown beyond what the company's own explanations provide. But much of the information obtained in the investigation is not public, and it is difficult for third parties to independently verify the conclusions. In this assessment, Anthropic had the authority to unilaterally remove portions containing non-public information, and METR explains that most of the evidence that was especially important in reaching its conclusions was removed.
The report organizes the information examined into seven categories and explains the disclosure status of each. Of these, information on internal capability benchmarks was disclosed, but analyses of AI's contribution and the content of researcher interviews were only partially disclosed. Details of intermediate R&D indicators, algorithmic progress, and major input resources are not public, and no specific information has been disclosed about bottlenecks that impede capability gains. An outside organization having checked internal information must therefore be distinguished from the underlying data being made public so that anyone can verify it.
There were also constraints on the access granted in the investigation. METR was not given the same privileges as Anthropic employees, and the environment did not allow it to directly check data, documents, and models without intermediaries. In addition, METR did not request, and did not receive, sampling to actually try unreleased models. This investigation was a separate matter from the plan for independent evaluators that Anthropic announced on October 15, and work to examine the allocation of compute between capability gains and safety measures was also dropped from the evaluation before conclusions were reached.
On the independence of the evaluation, METR explains that it conducted the investigation without receiving compensation. Anthropic did, however, provide free tokens and devices. In addition, the capability-gain assessment and the risk report review published this time are complementary documents produced by the same METR team as a single investigation, not two mutually independent follow-up studies. The fact that a third-party evaluation was carried out must therefore be considered separately from whether another third party could reproduce the same conclusions.
Regarding the outlook for the next 12 months, METR says it could not identify any clear constraint that would greatly impede capability gains in algorithmic progress, the compute used for training and experiments, data, and so on. However, its judgment changed repeatedly as additional information was obtained during the investigation, and its confidence in this outlook is low. There also remains the possibility that the acceleration rate is estimated as larger than it really is, due to excessive optimization on benchmark tasks with clear correct answers or to the low past growth rate used as the comparison baseline.
This assessment indicated that the pace of improvement in AI's R&D capability may have accelerated substantially compared with the past, while making clear that large uncertainty remains over how much of that acceleration can be explained as AI's own contribution. Going forward, it will be necessary to continue measuring the trajectory of capability gains under the same comparison conditions and to separate AI's contribution from factors such as hiring, compute, and algorithmic improvements. Whether the underlying data are disclosed in a form that third parties can verify will also be important. Only when such conditions are in place can the observation that "the pace of capability gains is about 2x" be tied to the risks posed by AI R&D automation and to judgments under the safety policy.