METR spent eleven weeks inside Anthropic and published the result on November 5. From August 12 to October 29, the Berkeley nonprofit was allowed to look at how fast Anthropic improves its models, and how much of that work the models now do themselves.
The answer: as of October 2026, Anthropic is improving its models’ capabilities on well-specified tasks at roughly twice the rate it managed between March 2024 and October 2025. METR attaches heavy uncertainty to that number itself. It would not be surprised by 1.3x or by 3x.
What METR got to see
Anthropic opened up data it had never shared: pretraining metrics, reinforcement learning metrics, training compute per domain, benchmark scores for internal models, its own analyses of the speedup from AI, forecasts for a future model, plus interviews with the people doing the work.
The hardest number comes from AECI, Anthropic’s version of the Epoch Capabilities Index. Since October 2025 the benchmark score there has been growing at 1.9x the earlier rate, with a 95 percent interval of 1.5 to 2.8. Over the past year, in other words, Anthropic covered about two years of its historical curve.
METR rules out compute as the explanation. The growth rate of training compute did not accelerate. Roughly 80 percent of today’s acceleration is attributed to AI itself, based on interviews, surveys and several lines of evidence, with substantial uncertainty. The work still runs mostly as assistance rather than autonomous contribution. METR does see signs that this is about to shift.
No brakes showed up. Neither algorithmic progress nor training compute nor experiment compute nor data looks like a limit that would stop further growth in the next twelve months. The report also notes that this picture changed several times during the assessment.
The most important evidence is redacted
Anthropic held the right to strike non-public information from the post, and METR held the right to say that it had. Both used those rights. The high-level conclusions about risk stand freely in the text. Of the evidence METR found most informative for those conclusions, the majority is redacted. Anthropic says the cuts remove confidential material that could accelerate other AI developers.
METR spells out what follows from that: the report does not achieve “transparency about evidence.” The public has to rely on METR’s judgment far more than is desirable for questions of this weight.
The timing is the real finding
The assessment ran during exactly the stretch in which Dario Amodei told the industry to slow down. His September 12 essay asked the field to “slow the pace at which we improve the capabilities of AI models.” In the same quarter, his company doubled that pace.
Anthropic did not immediately comment on the report. That these numbers are public at all is more than other labs allow, and it is precisely the half nobody outside can check.