METR: Anthropic's well-specified capabilities are improving ~1.9× faster since Oct 2025 — and most of that looks like AI-on-AI R&D

METR spent eleven weeks inside Anthropic with high access to internal R&D metrics, and the public write-up is blunt about both what it found and what it cannot show.

In Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, published November 5, METR's subjective conclusion is that as of October 2026 Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate it did between March 2024 and October 2025. The authors say they would not be surprised if the true figure sat closer to 1.3× or 3×. A majority of the current acceleration, they guess, is attributable to Anthropic's extensive use of AI for R&D — with substantial uncertainty, and with non-AI factors likely more important in late 2025.

The number that travels: 1.9× on AECI

The headline empirical finding is more precise, and narrower. Since October 2025, Anthropic's AECI scores — Anthropic's version of the Epoch Capabilities Index, a composite of many benchmarks — have been growing at an estimated 1.9× the pre-October 2025 rate, with a 95% confidence interval of 1.5–2.8×. METR glosses that as roughly two years of historical progress packed into the last year.

Two caveats sit right next to the number:

• AECI is about verifiable / well-specified tasks. METR is explicit that translation to messier, real-world work is an open question.
• The confidence interval only covers uncertainty in each model's AECI estimate. Methodological choices (which models to include, how to fit the break) can push the range wider; METR notes sensitivity analyses spanning about 1.1–2.9×.

METR is more confident about what the acceleration is not. It was not an acceleration in the scaling of training compute. Beyond that, Anthropic staff themselves did not agree on a single cause, and much of the more informative internal evidence was redacted.

Redactions are part of the story

Anthropic had unilateral rights to redact non-public information. METR says Anthropic redacted a majority of the evidence METR found most informative for its conclusions, describing the cuts as aimed at removing material that could accelerate other AI developers. Intermediate capability measures, algorithmic-progress estimates, most R&D-input metrics, detailed bottleneck analysis, and forecasts for future models are among the items METR saw but could not publish.

The public therefore has to lean harder on METR's judgment than METR would like. The authors say so themselves: the report does not achieve "transparency about evidence."

What still made it out is useful. Staff surveys of self-reported AI uplift show consistent growth and very high levels (METR cites reports reaching 12× in October 2026), though absolute uplift is not the same as an acceleration rate. R&D speedup today is described as mostly AI assistance rather than fully autonomous contributions, with "signs of life" that more autonomous work may arrive soon. METR did not see convincing bottlenecks in algorithmic progress, training compute, experiment compute, or data that would bind in the next 12 months — and immediately adds that its view of those bottlenecks changed several times as more detail arrived.

Companion review of Anthropic's August Risk Report

The same METR team also published a review of the "Risks from automated R&D" section of Anthropic's August 2026 Risk Report.

They agree with Anthropic that, for models available as of July 15, catastrophic risks from directly automating key R&D or from acute acceleration were low. They do not cleanly endorse Anthropic's claim that its early-warning threshold had not been crossed: that judgment, METR writes, rests on contestable calls that the public report does not communicate clearly. Both METR researchers and Anthropic staff lean toward "not crossed," but METR thinks crossing is plausible, and that the public report understates the acceleration that was already underway.

What a reader should take away

This is not a press release about a new Claude. It is an outside evaluator saying, with high-access data and heavy redactions, that Anthropic's measured capability trend on its own composite benchmark roughly doubled in slope after October 2025, that training-compute growth does not explain it, and that AI assistance inside the lab is the leading guess for most of the current effect — while warning that further automation could accelerate the accelerator.

Treat the 1.9× figure as a slope change on a lab-specific index of well-specified tasks, not as "AI is twice as intelligent." The redaction table is as important as the number.