METR, the independent evaluation group, says Anthropic is now improving its frontier models at roughly twice the pace it did a year earlier, and that AI tools are doing much of the work behind that speedup. The assessment, published on November 5, 2026, is based on about eleven weeks of access to Anthropic's internal data and roughly 15 interviews with its researchers. It is one of the most detailed outside looks yet at how AI is changing the process of building AI.
The Core Finding
METR's central estimate is that Anthropic improves its models on well-specified tasks at around twice the rate it did between March 2024 and October 2025. Its benchmark-based estimate puts the acceleration at 1.9 times after October 2025, with a 95 percent confidence interval of 1.5 to 2.8 times. METR says it would not be surprised if the true figure were closer to 1.3 or 3 times.
The group judges that about 80 percent of the current acceleration comes from AI, with substantial uncertainty around that number. METR is confident that faster growth in training compute was not the main driver. The speedup, in its reading, mostly reflects researchers using AI assistance rather than fully autonomous AI contributions to model training.
Self-Reports Versus Evidence
Anthropic staff answer informal Slack polls about how much the latest internal model speeds up their work. By October 2026, the average self-reported uplift reached about 12 times. METR calls that figure only very loose evidence, because the sample is unrepresentative and responses were often low-effort.
Interviews painted a more restrained picture, with views centered on roughly two times and a wide range around it. For anyone reading the headline numbers, that spread is a reminder that self-assessment of productivity gains is hard to trust.
The Baseline Problem
METR flags an important caveat. Before late 2025, Anthropic's models trailed the broader frontier on the benchmark measure, which makes its own historical baseline unusually low. When METR uses the models of all frontier developers as the reference, the estimated acceleration falls to about 1.4 times. Robustness checks across different statistical choices produced results ranging from 1.1 to 2.9 times.
In other words, the size of the speedup depends heavily on what you compare against. The direction of travel is clearer than the exact multiple.
Redactions and Missing Evidence
Transparency is another limit. Anthropic had the right to redact material before publication, and METR says it redacted a majority of the evidence it found most informative. METR writes that the public therefore has to rely on its judgment more than would be ideal. Anthropic has said the redactions aim to remove confidential information that could help other developers accelerate.
A planned audit of how Anthropic divides compute between capabilities work and safety research was also canceled mid-pilot, according to the reporting. That gap matters because compute allocation is one of the few concrete levers outsiders can use to check how a lab balances competing priorities.
Safety Thresholds and Outlook
The findings also touch on a risk threshold. Anthropic's August 2026 public risk report said the acceleration was less than a factor of two. METR reports that the internal, unredacted figure was 1.97, and that Anthropic described crossing the threshold as plausible. METR says the warning level may have been reached as early as July 15.
On bottlenecks, METR found no convincing constraint on algorithmic progress, training compute, experiment compute or data over the next 12 months. It also says its own view of the bottlenecks shifted several times, so it is not confident in that outlook. METR currently rates acute catastrophic risk as low, but it cannot rule out another sharp acceleration within six months.
What the Audit Did Not Settle
Several questions remain open. METR examined a specific benchmark composite, so the results describe progress on well-specified tasks and not necessarily on every capability that matters for safety or commercial value. The composite rose from about 15 points per year to 29, according to The Decoder's account, which is a large change but one measured on a particular yardstick.
The audit also had limits on access. METR reports that it was given access between August 12 and October 29, 2026, and says it was paid nothing by Anthropic. That independence is valuable, yet the redaction rights meant that Anthropic controlled which of the most informative findings reached the public. Readers should therefore weigh METR's conclusions as an expert judgment made under constraints, not as a fully reproducible measurement.
It is also unclear how the finding would look at other labs. METR notes that the industry-wide figure is about 1.4 times when models from other developers are included. If other companies are seeing similar benefits from AI-assisted research, the pace of the whole field may be increasing together.
Why It Matters
The idea of AI helping to build better AI has moved from speculation to measured claim, and it now sits inside a specific company's numbers. Whether the speedup is 1.4 times or nearly 3 times, the implications are significant for how quickly capabilities could advance and how quickly safety teams must respond.
For policymakers and enterprise buyers, the lesson is to demand independent, verifiable evidence rather than self-reports. For researchers, the work shows that outside evaluators can gain meaningful access to internal data, even if much of it stays confidential.
