I think it's just routing to faster hardware:

H100 SXM: 3.35 TB/s HBM3

GB200: 8 TB/s HBM3e

2.4x faster memory - which is exactly what they are saying the speedup is. I suspect they are just routing to GB200 (or TPU etc equivalents).

FWIW I did notice _sometimes_ recently Opus was very fast. I put it down to a bug in Claude Code's token counting, but perhaps it was actually just occasionally getting routed to GB200s.