METR says Anthropic is likely improving its models about twice as fast as before
Anthropic redacted a majority of the evidence METR found most informative, describing the redactions as removing information that could accelerate other developers, so readers must rely on METR's judgment more than METR considers ideal. METR says its team still disagrees internally. The 1.9x benchmark trend may be an artefact of the method or of overfitting, and it covers well-specified tasks, not all capabilities. Most relevant researchers at METR and Anthropic lean toward the threshold not having been crossed on July 15. The 1.97 is METR's calculation and does not appear in the private report. This covers one developer and is not a measure of real-world impact.
What was claimed?
METR, an independent evaluator, published on November 5, 2026 the results of a high-access assessment carried out at Anthropic from August 12 to October 29. Most METR staff on the project now believe Anthropic is likely improving its models on well-specified tasks at around twice the rate of March 2024 to October 2025, and METR would not be surprised by a true figure closer to 1.3x or 3x. Benchmark scores on Anthropic's internal index have grown at an estimated 1.9x the earlier rate since October 2025 (95% confidence interval 1.5x to 2.8x). METR guesses that perhaps 80% of the current acceleration comes from AI, and is confident it did not come from faster growth in training compute. In a companion review, METR agrees that catastrophic risks from Anthropic's July 15 models were low, but calls the public Risk Report's argument that the company's 2x early warning threshold had not been crossed overstated, and says a crossing by July 15 is plausible.
What does the evidence support?
Independent evidence, scope limited
If AI starts to speed up AI development, capabilities could improve faster than outsiders expect. METR describes its access as unprecedented, including internal data and staff interviews. METR found no clear bottleneck in the next 12 months and sees a significant risk of even faster acceleration, though it is not confident in that bottleneck assessment. The review also found a gap in public wording: the Risk Report said 'less than a factor of 2', while METR calculates that the private version implies 1.97.
What are the limits?
Anthropic redacted a majority of the evidence METR found most informative, describing the redactions as removing information that could accelerate other developers, so readers must rely on METR's judgment more than METR considers ideal. METR says its team still disagrees internally. The 1.9x benchmark trend may be an artefact of the method or of overfitting, and it covers well-specified tasks, not all capabilities. Most relevant researchers at METR and Anthropic lean toward the threshold not having been crossed on July 15. The 1.97 is METR's calculation and does not appear in the private report. This covers one developer and is not a measure of real-world impact.
Claim or evidence date: 2026-11-05. Last verified: .
METR · Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 ↗
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →