An outside review finds Anthropic's model progress has roughly doubled in pace, but most of its key evidence was redacted
METR spent eleven weeks with Anthropic's internal data and estimates its models are improving on well-specified tasks at about twice their earlier rate, mostly because of AI help. Staff self-reported a 12x speedup, which METR calls very loose evidence.
METR published an independent assessment on November 5 of how much AI is speeding up AI research and development at Anthropic. METR staff worked from August 12 to October 29, 2026, with access to internal data that included pretraining and reinforcement learning metrics, benchmark scores for internal models, lines of code and pull requests, AI agent inference spend over time, internal surveys of self-reported uplift and about 15 interviews with key staff. Its headline conclusion is that, as of October 2026, Anthropic "is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025." METR adds that "we would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate."
What the measured data shows
Since October 2025, Anthropic's scores on AECI, its version of the Epoch Capabilities Index, have grown at an estimated 1.9x the earlier rate, with a 95% confidence interval of 1.5x to 2.8x. METR says it is confident the speedup did not come from faster growth in training compute, and guesses that "a majority (perhaps ≈80%)" of the current acceleration is due to AI, a judgment it gives "with substantial uncertainty." Staff told METR that research is accelerated "primarily via AI assistance rather than fully autonomous contributions." The public version is limited: Anthropic "redacted a majority of the evidence we found most informative for these conclusions," which it described as "aimed at removing confidential information that could accelerate other AI developers," so "the public has to rely on METR's judgment much more than would be ideal." Anthropic also canceled a planned verification of its compute allocation midway through the project.
What staff self-reports can and cannot show
In Slack polls, staff put their average self-reported uplift from the most capable internal model at about 12x in October 2026. METR calls this "only very loose evidence," expects the answers came "from an unrepresentative subset of staff," and notes "some evidence that self-report survey estimates of uplift can overestimate true productivity impact." It also cautions that individual speedup estimates do not correspond to acceleration rates, so the 12x and the roughly 2x figures measure different things.
Why it matters
Even with unusually deep access, the evaluators leaned most on a measured benchmark trend, checked it against compute data, interviews and surveys, and still ended with a wide range. A business judging its own AI spend faces the same problem at smaller scale: how much faster people say they work is a reason to measure, not a measurement, and the case for the spend should rest on output records that show the change.