We ran real, authenticated streaming completions against xAI's grok-4.6 — with reasoning effort set to "low" — from Amsterdam, San Francisco, Montreal, Singapore and Tokyo, 20 per city, parsing the token stream itself and timestamping the moment the first visible token arrived.
Independent measurement by LatencyRadar. Not affiliated with xAI. · Last updated August 2026.
Grok answered all 100 requests, and where you call it from barely matters: the first visible token arrived in 2.08 seconds (Amsterdam) to 2.43 seconds (Singapore) — a 17% spread, the flattest regional profile we've measured for an LLM API. The stream itself opens fast, in 0.15–0.29 seconds, and grok-4.6 starts emitting reasoning deltas around 0.7 seconds in; the rest of the wait is the model thinking, now directly visible in the stream. Even at the "low" reasoning effort we requested, and for a trivial prompt, plan on roughly two seconds before the first visible word — and the slowest requests we observed took 2.9 to 4.8 seconds.
Amsterdam was fastest at 2.08 s to the first visible token, then Montreal (2.14 s), San Francisco (2.23 s), Tokyo (2.26 s) and Singapore (2.43 s). A 349 ms spread on a two-second wait means geography is nearly irrelevant for grok-4.6: when the model spends most of the time thinking, moving closer to the servers buys you very little.
Marker size and color reflect p95 latency by region.
Measured from the stream itself, one grok-4.6 call has three acts: the stream opens in 0.15–0.29 s (that part is network and edge), reasoning deltas start arriving around 0.7 s, and the first visible token lands at 2.1–2.4 s — with the trivial answer completing within about 20 ms after it. Unlike the ChatGPT API, which holds its response until the first token is ready, Grok's wait happens inside an open stream. The first visible word is what a user actually sees appear, so that's the headline number here — the transport phases below are supporting detail, and the long "download" phase is the in-stream reasoning, not bytes on the wire.
DNS resolution is measured once per region before the requests run, so it sits outside these bars: Amsterdam 2 ms · San Francisco 50 ms · Montreal 3 ms · Singapore 7 ms · Tokyo 5 ms
Phases come from the typical (median) request, so they add up to its total.
api.x.ai resolved to a single Cloudflare IP from all five cities, and our delivery verdict — computed from stream-open times rather than the inference-dominated totals — reads as anycast/edge with a 142 ms spread. First-visible-token times are nearly flat across cities, which says serving capacity, not network distance, sets the pace. This run's slowest observation was 4.8 s (Tokyo); an earlier run saw a single 12-second response from Amsterdam. One-off outliers happen, and so far they haven't followed geography — though single runs can't pin down where they originate.
All 5 regions resolved to one IP address and latency is uniform (142ms spread). Traffic appears to be served from an anycast or edge network.
Start with the headline number for each city. Open the technical breakdown only if you want to see where that time is coming from.
20 requests · 100% successful
20 requests · 100% successful
20 requests · 100% successful
20 requests · 100% successful
20 requests · 100% successful
The typical two-second first token is predictable — but the slowest requests we observed took 2.9 to 4.8 seconds, and an earlier run saw a single response take twelve, for the same trivial prompt at the same low reasoning effort. A closer region or a smaller prompt doesn't appear to fix that kind of outlier. If you build on a reasoning model, design for it: stream partial output as it arrives — with Grok you can even surface the reasoning phase as live progress — show that the request is in progress, and set timeouts generously enough that a slow-but-successful answer isn't killed at the finish line.
We ran real, authenticated streaming completions — POST /v1/chat/completions with stream: true — against xAI's grok-4.6 from servers in Amsterdam, San Francisco, Montreal, Singapore and Tokyo, 20 per city, all with the same fixed prompt. We parsed the SSE token stream and timestamped each event as it decoded: when the stream opened, when the first reasoning delta arrived, when the first visible token arrived (the headline number on this page), and when the response completed — alongside the usual DNS, connect and TLS phases. xAI opens the response stream immediately and reasons inside it, and the timestamps show it directly: headers in ~0.2 s, reasoning from ~0.7 s, first visible token at 2.1–2.4 s.
Run a free speed test from multiple cities and find out where your users are waiting — no setup, no account required.
Takes about 30 seconds.