METR Says Anthropic Is Likely Improving AI Twice as Fast but Redacted Key Evidence
An outside evaluator says Anthropic is likely improving its models about twice as fast as before, mostly thanks to AI itself, but admits Anthropic redacted most of the evidence behind that call.
An outside evaluator spent eleven weeks inside Anthropic asking one question: is AI now speeding up the work of building AI? The answer from METR, published November 5, is a cautious yes. The catch is that Anthropic redacted most of the evidence that got METR there.
In its independent assessment of AI R&D acceleration at Anthropic, METR says 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." Then it hedges hard, saying it "would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate."
That is not a claim of runaway self-improvement, and METR does not make one. It says it did not pay special attention to the thresholds in Anthropic's Responsible Scaling Policy, aiming instead for a "holistic picture" of AI-driven acceleration.
Two years of progress in one
The hardest number in the report comes from AECI, Anthropic's version of the Epoch Capabilities Index. METR says benchmark scores for Anthropic models, including unreleased internal ones, have grown at an estimated 1.9x the pre-October 2025 rate, with a 95% confidence interval of 1.5x to 2.8x. In METR's words, that implies Anthropic "has made roughly two years of historical progress over the last year."
That figure is higher than the 1.7x reported in Anthropic's own Haiku 5.5 system card. METR says the gap comes mainly from including data on newer models that have not been released. Different statistical assumptions swing the estimate from 1.1x to 2.9x, according to METR's robustness checks. When non-Anthropic frontier models are added to the chart, the measured acceleration drops to about 1.4x.
So what is driving it? METR says it is "confident" the speedup did not come from faster growth in training compute. Its best guess is that "perhaps" about 80% of the current acceleration is due to AI, a judgment it says carries "substantial uncertainty." Staff interviews pointed the same way. Anthropic employees voting in internal Slack polls put their average self-reported uplift from the most capable internal model at about 12x in October 2026, though METR calls those polls "only very loose evidence."
"It is hard to see how I am not being 3x more productive. I feel like I have postdocs and grad students, and we are working towards a larger project." Anonymous Anthropic researcher, quoted by METR
Not everyone bought it. Another researcher quoted in the report guessed that only "15-20% is due to AI" and said they "don't believe in self-reported productivity increase." METR also notes that AI cannot yet make highly significant, fully autonomous contributions to the most important parts of model training, although researchers described "signs of life."
The line that should make people sit up concerns limits. METR says it found no convincing bottlenecks in algorithmic progress, compute or data for the next 12 months, and warns of "a significant risk that further capabilities progress could lead to even faster acceleration." It also admits its view of those bottlenecks "changed several times," so it is "not confident" in that assessment.
Then there is the transparency problem, which METR raises itself. Anthropic had unilateral redaction rights over non-public information, and METR says Anthropic "redacted a majority of the evidence we found most informative" for its conclusions. Anthropic described the cuts as "aimed at removing confidential information that could accelerate other AI developers."
"As a consequence, the public has to rely on METR's judgment much more than would be ideal, given the weight of the questions answered here." METR
METR also discloses that Anthropic changed the project scope midway, dropping a planned check of how its compute is split between capabilities and safety before METR reached any substantive conclusions. Anthropic said it wanted to keep the engagement focused on risk assessment and would revisit compute verification later, according to the report.
The report lands days after Anthropic's own filing language about AI speedups, which this site covered last week. The difference this time is an outside set of eyes. Those eyes saw a lot. Readers, by design, get to see far less.
