METR's evaluation of Anthropic's internal models has forced a correction: the AI Futures Project now believes their 2027 predictions were five months too conservative, not too aggressive.

The revised forecast is specific and stark: AI matching the best human AI researchers arrives in 2027, and there is a 35 to 65% probability of systems capable of handling all white-collar work before 2028. These are not fringe numbers from optimists. These come from the forecasters who recently walked their timeline back toward the 2030s, then got pulled forward again by new internal benchmark data.

The piece is worth reading in full because the argument is not built on hype. It is built on a named report, named organizations, and named researchers who say they are afraid of what they are measuring. The gap between public AI products and what labs are testing internally is the actual story here.

[WATCH ON YOUTUBE →]