An independent evaluation of Anthropic’s internal research and development data estimates that benchmark capabilities on well-specified tasks have advanced about 1.9 times faster since October 2025 than during the preceding period. The finding, published Nov. 5 by the nonprofit Model Evaluation and Threat Research (METR), is an estimate with a wide 95% confidence interval of 1.5 to 2.8 times—not evidence that AI development has escaped human control or that the same rate applies to every kind of work.
What METR measured
METR’s report, dated Nov. 5, describes an assessment conducted at Anthropic from Aug. 12 through Oct. 29, 2026. The researchers examined whether the pace of capability improvement at the company had accelerated and how much of any change could be attributed to AI assistance in research and development. Their conclusion is specific: by October 2026, most METR staff working on the project judged that Anthropic was improving models on well-specified tasks at roughly twice the rate seen between March 2024 and October 2025. The team said it would not be surprised if the actual acceleration were closer to 1.3 times or 3 times. [METR assessment]
The central numerical result comes from benchmark scores plotted using AECI, Anthropic’s version of the Epoch Capabilities Index. METR estimates that AECI scores for Anthropic models have grown 1.9 times faster since October 2025 than before then, with a 95% confidence interval from 1.5 to 2.8. Put in the report’s plain-language framing, models appeared to make about as much benchmark progress in the latest year as in the previous two years. This describes a rate comparison on a selected index; it is not a claim that models’ general competence doubled in every domain.
METR uses “acceleration” to mean a ratio between rates of capability growth. A twofold rate means that one year produces what would previously have taken two years, according to the selected measure. It does not mean that a system has doubled its intelligence, become twice as autonomous, or acquired a general ability to improve itself. METR explains this terminology because familiar words such as “twice as fast” can otherwise blur a benchmark trend with a much broader claim.
The reported 1.9 figure is a fitted trend, not a direct stopwatch measurement of every part of Anthropic’s work. The report says growth rates depend on decisions such as which benchmarks and models to include and how to date model snapshots. Robustness checks under alternative choices yielded acceleration estimates from 1.1 to 2.9 times. METR says it would be surprised if a better version of AECI showed no acceleration or more than threefold acceleration, but that judgment itself reflects the researchers’ analysis of a sparse and evolving dataset.
An unusual high-access review
Unlike an assessment based only on public model releases, METR says Anthropic made internal information available. The report lists measures relating to pretraining, reinforcement learning, compute used in different domains, internal-model benchmark scores, assessments of AI speedups, projections of future-model metrics and staff interviews. The independent review group describes the access as unprecedented. AP independently reported that Anthropic provided internal data over roughly 11 weeks and that METR interviewed employees. [Associated Press coverage]
That access strengthens the assessment compared with an outside observer attempting to infer research speed from public product launches alone. It also creates a transparency problem. Anthropic could redact nonpublic information from the public report, and METR says the company removed a majority of the evidence the assessors found most informative for their conclusions. Anthropic characterized the redactions as intended to prevent confidential details from accelerating other AI developers. METR says it was permitted to disclose the existence and importance of redactions, but some remaining detail could not be described without revealing protected information.
Consequently, readers cannot independently inspect much of the evidence underlying the headline conclusion. METR says the redacted material had multiple plausible interpretations and that both its own researchers and Anthropic staff had substantial uncertainty about how to combine it. The group explicitly acknowledges that the public must rely more heavily than ideal on METR’s judgment. This is not a minor footnote: the 1.9-times benchmark result is visible, but the wider case for how AI contributed to the company’s overall progress is partly based on information unavailable for external scrutiny.
Why METR attributes much of the change to AI
METR considered multiple possible explanations for faster progress, including training compute, hiring, data improvements, algorithmic discoveries and the removal of other constraints. It says it could confidently rule out an acceleration in the growth rate of training compute as the cause. The report does not say compute is irrelevant to model performance; rather, the rate at which training compute scaled did not explain the observed change in the pace of capability gains.
On the contribution of AI, METR offers a subjective attribution rather than a clean causal measurement. Its researchers guess that a majority—perhaps about 80%—of the acceleration as of October 2026 is due to AI use in R&D. METR stresses that this estimate has substantial uncertainty and says non-AI factors likely played a significant role in the acceleration that emerged in late 2025. It reports no decisive evidence or consensus among Anthropic staff on all possible causes. The 80% figure should therefore be read as the evaluators’ best judgment, not an audited accounting of research output or a precise measured share.
Several strands of evidence informed that judgment. Interviews with researchers produced a wide range of estimates but, in METR’s overall impression, centered near a twofold speedup in their areas of work. The report also describes staff surveys and self-reported uplift from using AI tools. METR calls the survey evidence loose: quick self-estimates are difficult to interpret and do not by themselves establish how much useful research was produced or how much faster a project reached a result.
The account that METR received was that AI primarily assists researchers rather than independently carrying out the most important components of model training. The evaluators nevertheless saw indications that more autonomous contributions could become more important. AP likewise reports the distinction: the assessment did not find that AI was already making highly significant, fully autonomous contributions to the most important parts of training, while describing early signs that the role could expand. Neither source establishes that a recursive self-improvement loop is already operating independently of human direction.
METR also distinguishes the late-2025 shift from the current attribution. It believes non-AI factors likely contributed to the initial acceleration. Its view that AI’s relative contribution grew over time rests on several lines of evidence, some redacted, and on its judgment that other investigated factors do not explain a sustained higher trend. This is an interpretation of a complex research process, not a controlled experiment in which AI access was switched on for one team and off for another.
The limits behind the headline number
AECI covers well-specified tasks: problems whose success can be evaluated comparatively clearly, including through automatic grading or language-model assessment. Such tasks are useful for consistent comparisons, but they are not interchangeable with every activity involved in research, engineering or deployment. Real projects can involve ambiguous goals, iterative judgment, coordination and consequences that are hard to compress into benchmark scores. METR says there is additional uncertainty about how the measured acceleration translates to broader, less well-specified work.
The index itself could contribute to the observed pattern. METR warns that acceleration could reflect overfitting to the benchmark distribution or an artifact of AECI’s methodology. The benchmark observations are relatively sparse and lag developments in research, so the index may not reliably capture the timing of a recent change. The report includes unreleased internal model snapshots in its analysis; its estimate therefore is not simply a public leaderboard comparison that readers can reproduce from product announcements.
The comparison baseline also matters. METR’s headline risk judgment compares the current pace with March 2024 to October 2025. Its empirical chart describes the AECI trend before and after October 2025, while its explanatory framing compares roughly the latest year with the prior two-year period. The report cautions that Anthropic’s early 2023–2025 growth may have been unusually low. If the earlier pace was atypical, the ratio can make today’s change look larger than comparison with a different historical window would.
There are uncertainties in both directions. METR says it could be underestimating acceleration if some growth rates have recently increased further, if AI tasks become more automated, or if benchmark growth understates effects in real-world use. It could be overestimating it if non-AI contributions mattered more than the assessors judged, if the index’s emphasis on verifiable tasks is unrepresentative, or if the baseline period was unusually slow. The assessment is not a forecast that acceleration must continue.
The report also discusses potential bottlenecks. METR did not identify clear near-term constraints from algorithmic progress, training compute, experiment compute or data that would prevent further capability progress over the next year. But it says its understanding of these bottlenecks changed as it received more detail and that it is not confident in the assessment. “No bottleneck observed” is not equivalent to proof that none exists; hidden constraints, organizational limits or a shift in the underlying research problem could matter.
What the assessment does—and does not—signal
METR frames the possibility of compounding AI-assisted research as a safety concern: if stronger systems help accelerate development of stronger successors, progress could potentially speed up further. It says the lack of observed bottlenecks raises significant risk that capability progress could yield faster acceleration over the coming year. That is a risk judgment, not a prediction that runaway development will happen, nor a claim that current systems have crossed a threshold into general intelligence or escaped oversight.
The distinction matters because the report combines three different kinds of statements. The AECI estimate is a quantitative trend analysis. The estimate that AI accounts for a majority, perhaps around 80%, of the current acceleration is a subjective causal judgment. The concern that acceleration could compound is a forward-looking risk assessment. Each has a different evidentiary footing, and collapsing them into one assertion would make the findings sound more certain than METR says they are.
AP’s independent account corroborates the date, the high-access process, the headline conclusion and the caveats about narrow tasks and redactions. It also places the evaluation in the context of public debate about AI development and safety. The AP story is corroboration of the report’s publication and central interpretation, not a second independent measurement of Anthropic’s internal benchmark data; its key numerical and methodological evidence ultimately comes from METR’s assessment.
For now, the defensible takeaway is bounded: METR reviewed substantial internal information and estimates that Anthropic’s measured progress on a particular class of tasks accelerated, with benchmark trends consistent with roughly a twofold rate increase. METR believes AI assistance probably explains much of the present acceleration, while emphasizing that the evidence is incomplete, some of the most informative details are withheld, non-AI factors mattered, and the result may not generalize beyond the index. It is a notable assessment of AI’s role in AI research, not proof of runaway AI or a general measure of intelligence.
Frequently asked questions
What does the 1.9-times figure mean?
It is METR’s estimate that Anthropic’s AECI benchmark scores grew about 1.9 times faster after October 2025 than before, with a 95% confidence interval of 1.5–2.8. It compares rates of benchmark progress, not model intelligence or performance across all tasks.
Did METR conclude that AI caused all the acceleration?
No. METR’s subjective estimate is that AI accounts for a majority—perhaps about 80%—of the current acceleration, with substantial uncertainty. It says non-AI factors likely contributed significantly to the late-2025 acceleration.
Why can’t readers fully check the evidence?
Anthropic exercised redaction rights over much of the evidence METR considered most informative. METR says the redacted information is difficult to describe further without disclosing sensitive material, limiting public scrutiny of its broader judgment.
Does the report prove recursive self-improvement is happening?
No. METR says AI assistance is primarily augmenting researchers rather than making fully autonomous contributions to the most important training work. It flags signs that autonomous contributions may grow, but does not claim that an uncontrolled self-improvement process is underway.
Can the AECI result be generalized to ordinary work or general intelligence?
Not from this report alone. AECI focuses on well-specified tasks that are easier to evaluate, and METR says it has additional uncertainty in translating the result to broader, less clearly defined tasks.