# Methodology — how AI API latency is measured
Updated July 24, 2026. Latency is measured directly from probes in four regions (Europe, US, Asia, South America) — not scraped, not through a gateway.

## What is measured
- Edge latency (time-to-first-byte, TTFB): DNS + TCP + TLS + first response byte to the provider's official API host.
- Inference latency (TTFT): needs an API key per provider, so measured only where we hold one and shown in a separate table. Headline rankings are always edge latency (TTFB), never mixed.

## Where from
- Europe (Germany), US (Central), Asia (Tokyo), South America (São Paulo). Rankings differ by region.

## How often
- Repeating schedule (every few minutes); figures are p50/p95 over the last 24h with a visible update date.

## Percentiles
- p50 = median (typical request); p95 = slow tail; uptime = share of successful probes. Nearest-rank percentiles; averages avoided.

## Speed Index
- Composite 0-100: per region, fastest_p50 / provider_p50 * 100 (100 = as fast as the region leader), averaged over measured regions. Homepage lists only providers measured in all regions. Reproducible, derived from measured p50.

## Scope
- Measures response latency and uptime by region over time. Not a benchmark of model quality, throughput or price. Every number is measured, nothing estimated.

Source: https://llmlatency.dev/methodology