METR puts Claude Opus 4.5 at a 4h49m 50% autonomy horizon, matching the fast variant of AI 2027 in which each doubling gets 15% easier. Prediction markets missed it, which the issue treats as the finding: the forecasting community is recalibrating for recursive self-improvement after the fact.
Underneath, recursion crosses into matter. CATL runs humanoids on the lines that build the batteries that power them.