TTFB and TTFT Measure Different Things — Mixing Them Distorts AI API Rankings

A developer running llmlatency.dev has published findings from continuous latency measurements across roughly 45 AI inference providers, tested from four global regions including Germany, the US, Tokyo, and São Paulo. The analysis draws a sharp distinction between TTFB (time to first byte), which reflects network and infrastructure speed before any model is involved, and TTFT (time to first token), which also includes provider queueing and model prefill time. Data shows that network latency accounts for only 3–33% of total time to first token, meaning the bulk of what most benchmarks measure is actually model behavior rather than API performance. The study also highlights that regional variation can be dramatic, with one provider showing an 18.9x difference in TTFB depending on the continent a request originates from. The author cautions that leaderboards comparing TTFT across providers without disclosing the underlying models are effectively ranking models against each other, not the APIs themselves.