Observe or expect

Twice the pace. Roughly 80% of it attributable to AI. A 12x uplift, by Anthropic staff's own informal reports. Those are the numbers in Ajeya Cotra's reading of METR's public report on Anthropic, which drew 328 points and nearly 300 comments on Hacker News on Friday. Between October 2025 and October 2026, capabilities progress at Anthropic ran about double the pace of the twelve months before, and AI doing the research is most of why it is still running that fast.

The numbers matter because of a line Anthropic drew for itself. Its Responsible Scaling Policy, in the version effective since July, names a threshold for "dramatically accelerated progress due to AI." In its S-1 this week, Cotra notes, Anthropic told prospective investors it "may imminently meet" that threshold. Cotra works at METR but didn't work on the report, and she puts it more plainly: "on a plain reading, my best guess is that it has been met."

The gap between those two sentences looks like a dispute over a measurement. It isn't. The policy settled it in advance. Here is the threshold, as Cotra quotes it:

Observe or expect. The second verb carries the weight. A doubling in the rate of progress can only be confirmed in arrears: by the time a year of data shows it, you have spent that year inside it. Anthropic wrote its trigger so that it fires on a forecast, not a confirmation. "May imminently" is not quite "expect," and the space between them is the room a securities lawyer leaves. But it is a forecast, and the policy was written to act on forecasts.

That is not a quibble over wording, because a threshold is a promise about what happens next. Anthropic's May 2025 policy attached a floor to a weaker milestone in the same family: security good enough to protect "against model-weight theft by state-level adversaries." Security of that grade is not something a lab bolts on in a quarter, and Cotra's assessment of where the industry stands is blunt: "both OpenAI and Anthropic recently had major security breaches by amateur groups." A trigger that waits for confirmation fires after the safeguards were needed. That is the reason the policy says expect.

A trigger that waits for confirmation fires after the safeguards were needed.

The best case for the S-1's wording is a fair one. The measurement is uncertain. METR's authors attribute about 80% of the speedup to AI, which puts the counterfactual figure near 1.8x rather than 2x, and they say their picture changed several times. A majority of the evidence they found most informative is redacted. A company about to start its IPO roadshow next week can't declare a safety threshold crossed on a guess, and calling it early would cheapen the line for the next time. All true. But uncertainty is the exact condition the "or expect" clause was written for. It moves the call off the measurement and onto judgment, and Anthropic has already given its judgment, in a filing its lawyers read line by line: imminently.

Next week that filing goes on the road. Investors will hear "may imminently" as a business risk, a thing that could slow the company down. The policy asks for the other reading: a line you expect to cross is a line you have crossed. Observe or expect. The second word was the safeguard.