METR: AI Has Doubled the Pace of Research at Anthropic — RSP Threshold Appears to Be Crossed

What Happened

This week Anthropic stated in its S1 filing that it "may be about to" reach the threshold of "sharply accelerated progress due to AI." The threshold is described in the Responsible Scaling Policy, in the current version from July 2026. There are two conditions. First: overall capability progress is running twice as fast as the fastest sustained pace seen before AI began to meaningfully contribute to research. Second: the acceleration is plausibly explained by the automation of AI R&D, rather than by growth in compute or other resources.

In parallel, METR staff spent several months running an investigation with deep access to Anthropic's data. Their report came out on November 5. Ajeya Cotra, having read it alongside the RSP, concludes that the threshold has, by all appearances, already been crossed.

What METR Found

The report's authors reached two conclusions.

First: from October 2025 to October 2026, capability progress at Anthropic ran roughly twice as fast as in the previous twelve months. It was measured with an internal version of the Epoch Capabilities Index, called AECI. The remaining supporting data is redacted. According to the authors, this is "most of the evidence we found most informative."

Second: the pace may have started to rise for various reasons, but it is now sustained mainly by the automation of research with the help of AI. The arguments for this are also blacked out.

There is a caveat as well: we see only the conclusion and a single public metric. It cannot be verified independently, so we have to trust METR and Anthropic itself.

Why It Matters

Cotra calls the report the most important update on AI risk in a couple of years, more important than the Hugging Face attack story. A direct consequence: timelines are probably shorter than outside researchers assumed, including the AI 2027 team.

The key piece of arithmetic is that if AI doubles the overall pace of progress, it speeds up human researchers far more. Some of the work still depends on compute, experiments, and other bottlenecks that AI does not accelerate. For the whole to grow twofold, the returns on the non-bottlenecked parts must be much greater than twofold.

What Practitioners Should Do

Developers and product managers should build a faster model-update cycle into their plans. If the leaders are truly accelerating, the API and agent capabilities you rely on will change faster than usual. Architectures tied to the weaknesses of a specific model will become obsolete sooner.

Founders should reconsider the horizon over which their product remains unique. If your advantage rests on something the model can't do yet, that cushion may shrink.

For those following safety: the RSP threshold is tied to the company's commitments on protective measures. Watch how Anthropic responds formally, since security requirements and conditions of access to models depend on it.

Finally, a word of caution. This is one author's interpretation of a partially redacted report. The full text by Cotra and the METR report itself are worth reading before building a strategy on this.