METR estimates Anthropic's capability progress has roughly doubled in pace, with most evidence redacted

METR, a Berkeley-based nonprofit that evaluates frontier models, published on November 5 an independent assessment of whether AI is speeding up research and development at Anthropic. A team equivalent to 2.5 full-time staff worked with the company from August 12 to October 29, reviewing internal benchmarks, compute and staffing data, surveys and about 15 staff interviews. METR says it took no payment but received free tokens and worked on Anthropic devices; it did not sample from unreleased models.

The headline estimate

METR concludes that, as of October 2026, Anthropic is likely improving capabilities on well-specified tasks about twice as fast as it did between March 2024 and October 2025. It adds that it would not be surprised by a true value nearer 1.3x or 3x, and that the link to broader, messier work is less certain.

Much of this rests on AECI, Anthropic's internal version of the Epoch Capabilities Index. METR estimates its growth since October 2025 at 1.9 times the earlier rate, with a 95% confidence interval of 1.5x to 2.8x; robustness checks span 1.1x to 2.9x. Including models from all frontier developers yields a lower 1.4x.

On causes, METR says it is confident the speed-up did not come from faster growth in training compute. Its best guess is that about 80% of the acceleration is due to AI, though it stresses large uncertainty. It found no convincing bottlenecks for the next year, but holds that view with low confidence.

Redactions and methodological limits

Anthropic had unilateral rights to redact the report, and METR says the company removed a majority of the evidence it considered most informative, including detailed attribution and forecasts. Anthropic described the cuts as protecting information that could help competitors. METR acknowledges that readers must rely on its judgment more than is ideal, that data came largely through two Anthropic staff, that Claude was heavily involved in collecting and analysing it, and that a planned check of compute allocation was cancelled by Anthropic midway. It also lists ways it could be overestimating, such as gains limited to automatically graded tasks or unusually slow growth before late 2025.

According to the Associated Press, via Fast Company, METR found that Anthropic's public August risk report understated the acceleration under way; METR says its review of that report will be published separately. Anthropic did not immediately comment, AP reported.