LLM API Latency Benchmark

Where it is measured from: 2 locations, reported separately

🇭🇰 Hong Kong, 🇺🇸 US Central (Chicago). The two columns are measured independently and never averaged together. The same endpoint routinely differs several-fold between the two, and an average of those would be a number nobody could use. The Hong Kong probe measures international and cross-border routes, which is not the same as a connection originating inside mainland China. Calls to Chinese models from the US probe cross the Pacific. There is no mainland probe yet; if one is added it will be labelled separately.

How it is measured

TTFT uses a streaming request with max_tokens between 8 and 16, recording when the first token arrives. Output speed caps generation at 256 tokens and divides the provider’s own completion_tokens by elapsed time. Reasoning mode is explicitly disabled for both, because chain-of-thought tokens count toward completion_tokens and ignore max_tokens (one model returned 896 tokens for a 256 token request). Leaving it on would compare thinking speed against answering speed. The cost is coverage: Gemini 3.x refuses to disable reasoning across the whole family, so this board uses Gemini 2.5 Flash instead. Newer flagships increasingly force reasoning on, so this comparable set will narrow rather than widen.

For rolling 7 and 30 day p50 / p95 figures with sample counts, see the latency dataset (JSON and CSV downloads included).

What this table does not do

It does not rank which provider is best. Latency is one dimension of a choice, and samples from two locations cannot represent worldwide experience. Sample counts are in the table, so you can judge for yourself how much weight a row carries.