On November 5, 2026, the independent evaluation group METR published a report on how much AI is speeding up Anthropic's own AI research and development. Its 4 primary authors ran an assessment of about 11 weeks, from August 12 to October 29, with permission to view some internal material. The work took about 2.5 full-time equivalents and covered pretraining, reinforcement learning, training compute, internal model scores and staff interviews. METR delivered a Full Report of about 100 pages to Anthropic. What the public sees is a redacted public version of it.
This piece does not argue over whether twice is too high or too low. It asks a different question: how far apart are an independent evaluator's conclusions and the evidence the public can actually see? METR answers that question itself at the top of its report, and quite directly.
The conclusions could be published, but much of the evidence METR leaned on could not
The public version opens with a note on redactions. Under the contract, Anthropic could redact any non-public information, and METR kept the right to say whether Anthropic used that right and whether any findings depended significantly on redacted material. METR writes that it could freely share its top-level conclusions about risk, but that "Anthropic redacted a majority of the evidence we found most informative for these conclusions."
That sentence needs a careful reading. It does not say most of the report was blacked out, and it does not say METR lacked evidence. METR saw the evidence and based its judgment on it; what was removed was the right to publish the most informative part, not the evidence itself. METR also says most of the redacted information was very high-level.
The report includes a table that sorts the evidence into 7 types, showing for each whether it was shared with METR and whether its results are described in the public version. Benchmark scores of internal models: shared, and described. Key intermediate measures and algorithmic progress estimates: shared, but not described. Measurements of key R&D inputs: mostly shared, not described. Details of bottlenecks and forecasts: partly shared, not described. Evidence of AI's contribution and staff interviews: shared, partly described. By that table, 1 of the 7 types has its results fully described in public.
METR does not play down the cost. It writes that the report does not achieve "transparency about evidence": the public will find it hard to judge the evidence itself or reconcile it with other sources, and so has to rely on METR's judgment much more than would be ideal.
Roughly twice is a subjective judgment with a wide range, and its public part is mainly one benchmark trend
The report's central judgment is that, as of October 2026, Anthropic is likely improving its models' capabilities on well-specified tasks at around twice the rate it did between March 2024 and October 2025. METR stresses the uncertainty: it would not be surprised if the true figure were closer to 1.3x or 3x, and it adds further uncertainty about how this carries over to broader, less well-specified tasks. This is the view of most of the METR staff on the project, not all of them.
Several lines of evidence feed that judgment, and the one written up most fully in public is this: measured by AECI, Anthropic's version of the Epoch Capabilities Index, Anthropic's models have been improving since October 2025 at 1.9x their earlier rate, with a 95% confidence interval of 1.5x to 2.8x. Put more plainly, the past year of benchmark progress matches the two years from October 2023 to October 2025.
This part was barely touched. METR says the section was very minimally redacted; only the names of specific benchmarks and one graph were removed, and the methods and robustness checks are in the public version. The robustness checks are worth reading: under different assumptions the acceleration ranges from 1.1x to 2.9x, and the most sensitive assumption is how task difficulty is distributed within benchmarks. When other companies' frontier models are included, the estimated acceleration falls to 1.4x. METR offers two explanations without settling on either: one is that Anthropic progressed more slowly than the industry frontier before late 2025, so a low baseline inflates the ratio; the other is that earlier models from others were more benchmark-optimized, which would make the historical trend look too steep.
METR also says AECI alone could overstate the acceleration or be an artifact of the method, for example by overfitting to the benchmark distribution. What made it take the roughly twice figure more seriously, and hinted that acceleration could be above twice and rising, was another body of evidence, and that evidence was redacted. In other words, the public can review the published methods, results and robustness checks, but that is not the same as being able to recompute them independently; and this public trend is not the evidence that most moved METR toward twice.
Attributing most of it to AI is an estimate, not a proven cause
The report's second judgment is that a majority of the current acceleration is likely due to Anthropic's heavy use of AI in its research. METR's guess is about 80%, given with "substantial uncertainty." It also says the acceleration in late 2025 was likely influenced by non-AI factors, and that details of many of those factors were redacted.
One point METR says it is confident about is that the acceleration was not caused by faster growth in training compute. But the details behind that statement were redacted too, so the public sees the conclusion. On other possible causes, such as hiring, better data, algorithmic discoveries or resolving other bottlenecks, Anthropic's own staff had no consensus, and METR says it found no decisive evidence.
There are two kinds of public supporting evidence. One is self-reported uplift: in Slack polls, staff estimated their speedup from the internal model in use at the time, and in October 2026 the average was about 12x. METR calls this only very loose evidence, since respondents may not be representative and answers reflect little effort, and it notes that an individual's speedup is not the same as an acceleration in overall progress. The other is interviews: METR had Claude classify interview quotes by whether they support an acceleration of twice or more. Going by the published interview classification results, METR reads them as "mildly in favor" of twice or more. The quotes themselves had sensitive details redacted and the sorting was done by a model, so this counts as supporting evidence only.
This material shows that people inside Anthropic broadly feel AI is speeding up their work, but it is still some distance from the causal claim that AI produced a doubling. In its limitations section METR lists the possibility in the other direction too: non-AI factors might explain most of the acceleration.
Why the evidence stays confidential, and what an audit adds
Anthropic described the redactions as "aimed at removing confidential information that could accelerate other AI developers." METR agrees the redactions were limited to that kind of information but declines to say whether they were justified overall, because it does not think it is well placed to weigh a company's interest in keeping such information confidential against the public's interest in understanding AI risk. Anthropic made a similar point in its announcement on independent evaluators, which METR's report dates to October 15: some safety-related information cannot be made public because it could help malicious actors or is commercially sensitive, such as compute or training information, so instead third-party evaluators get deep access and publish evidence and reasoned conclusions.
Within that arrangement, the audit offers something real. METR saw internal model data outsiders cannot see, including 4 internal-only models. It interviewed around 15 other Anthropic staff, many of them chosen by METR rather than Anthropic, and says it noticed no discrepancies between what they said and what its primary contacts said. METR took no payment; no fees went in either direction, and Anthropic provided only laptops and free token credits. The report also discloses a conflict of interest: one team member had been in talks with Anthropic about a role at the start, received and declined an offer, and held equity in a competitor as a former employee, so all of that person's work was checked by an unconflicted colleague.
METR also sets out the audit's limits. Its access was not "employee-like": it could not read documents on demand or observe internal discussion, a small number of data requests were denied, and in a small number of cases Anthropic withheld information it considered too sensitive. Over the assessment METR changed its reading of key data several times, in some cases reversing itself more than once on critical questions. Most of its data came through two Anthropic staff members, and Claude was involved in most of the data and analysis Anthropic provided and in most of METR's own analysis. The contract originally included verifying how Anthropic splits compute between capabilities and safety; Anthropic removed that component midway, before METR reached any substantive conclusions on it. METR also says it weighed how the report might affect future work with companies, and that Anthropic gave feedback on clarity, mainly on METR's review of its risk report, some of which METR felt was "overly forceful."
METR's overall view is that it is "extremely unlikely" Anthropic was actively trying to mislead it, but that greater access would make the assessment harder to mislead.
What this kind of evaluation is worth depends on how well its redactions are explained
Put these together and you can see what the public has. What it can review: AECI's methods, results and robustness checks (though that is not the same as recomputing them), loose evidence such as self-reported uplift and interview classifications, and METR's disclosures about process, contract and conflicts of interest. What is absent from the public version and depends on METR's judgment: the internal metrics that made it take roughly twice more seriously, the split between AI and non-AI contributions, the bottleneck analysis and the forecasts for future models.
None of this makes the conclusion untrustworthy or means evidence is missing. The evidence exists, people have examined it, and those people have said publicly what they saw, what they could not publish, and how much their conclusions rest on what they could not publish. That is what sets this report apart from an ordinary third-party endorsement: it does not ask readers to accept the conclusion unconditionally, but marks out where trust is required.
For readers, the safer approach is to keep two things apart. The 1.9x benchmark acceleration is a measurement with published methods and results, an interval and stated limitations. Roughly twice overall, and about 80% due to AI, are METR's subjective judgments, drawing on both public and redacted evidence, with wide ranges and disagreement within the team. How much weight the latter deserve depends on how much you trust the evaluator's independence and judgment; this time, METR at least put as much of the material needed to make that call in plain view as it could.