METR finds AI is accelerating Anthropic’s R&D roughly twofold
On November 5, 2026, the evaluation nonprofit METR published an 11-week audit of Anthropic and estimates its models are now improving 1.9 times faster than before October 2025, an acceleration mostly attributed to AI use in its own research. For teams that build or evaluate AI, it is the first independent evidence of a self-improvement loop at a frontier lab.
November 5, 2026. METR, the independent nonprofit that evaluates the most advanced AI systems, publishes an audit of Anthropic’s research and development. August 12 to October 29, 2026. The window of the high-access evaluation, run inside the lab. 1.9x. The measured acceleration in the pace of model progress, with a confidence interval of 1.5 to 2.8x. Why it matters: this is the first independent, quantified, sourced evidence that AI is speeding up the development of AI itself — a loop the field has long feared and no one had yet measured at this scale.
A high-access audit, a cautious conclusion
METR was given access to internal data Anthropic never shows: pretraining metrics, reinforcement learning, training compute by domain, benchmark scores for internal models, AI speedup analyses, and forecasts of future models. The report was read and redacted by Anthropic before publication.
The result is a rare exercise in honesty. METR states that the acceleration is real, but admits high uncertainty: it “would not be surprised” if the true figure were closer to 1.3x or 3x. And it flags that a majority of the most informative evidence was redacted, forcing the public to lean on METR’s judgment more than it would like.
1.9x faster, and it is not the compute
The central number is this: since October 2025, benchmark scores for Anthropic’s models — measured via AECI, its version of Epoch’s capabilities index — have grown at an estimated 1.9x the previous rate, with a 95% confidence interval of 1.5 to 2.8x. In plain terms, Anthropic appears to have produced about two years of historical progress in the last year.
What matters is what does not explain the acceleration. METR is “confident” on one point: it did not come from faster scaling of training compute. Interviews and surveys instead point to heavy use of AI in R&D itself. METR’s overall estimate: roughly 80% of the current acceleration is attributable to AI, with “substantial uncertainty.”
Bottlenecks not yet in sight
The report sets aside, for now, the possibility of a fast plateau. METR saw “no evidence” of major bottlenecks — algorithmic progress, training compute, experiment compute, or data — that would stop progress within the next 12 months.
But it warns immediately: its understanding of the bottleneck situation changed several times as more information arrived, and it is not confident in its current assessment. The absence of a visible ceiling is an observation, not a prediction. And because the most decisive evidence was redacted, the margin of error remains wide.
What this changes for evaluation and safety
The conclusion lands directly for anyone who evaluates or monitors AI. If AI is speeding up its own design, new capabilities can appear faster than we learn to measure them. That is precisely the dynamic safety teams fear: evaluations risk going stale faster than they are produced.
The finding echoes the scenario METR laid out beforehand: accelerated R&D could lead to the sudden emergence of extreme and risky capabilities before guardrails are in place. The Anthropic audit is the first to attach a concrete number to that hypothesis.
Redaction adds a second problem. METR says it plainly: by removing most of the evidence, Anthropic prevented the public from “judging for itself.” Transparency about data is becoming as strategic as transparency about results. For a CISO or a regulator, that is a signal: safety conclusions now rest more on trust in a third party than on checkable facts.
Verdict
The METR report does not announce a catastrophe — it documents an inflection. For the first time, an independent body quantifies an acceleration in a frontier lab’s R&D, attributes most of it to AI, and rules out compute scaling as the sole explanation.
If you evaluate or audit AI systems, treat this report as an anchor: the pace you calibrated against a year ago is probably obsolete, and your evaluations need to be rerun more often. If you follow AI safety from the outside, treat the redaction as a fact in itself — society’s ability to judge progress depends on evidence that, for now, remains partly invisible. The acceleration is real, it is measurable, and it is not waiting for evaluation to catch up.