That's for request handling overhead. Real-world speedup depends on your bottleneck (DB, external APIs, etc.), but the gains still apply. Good feedback though, and we have more benchmarks planned regardless.