Which LLM API has the lowest time to first token?

Measured over the last 24 hours, the lowest median time to first token (TTFT) is 757 ms — cerebras running gpt-oss-120b, requested from South America (São Paulo). Read that as one measurement, not as a provider ranking: every provider here answers on a different model, and model size moves TTFT far more than the provider does. What is comparable like-for-like is the same provider on the same model from different regions — and that alone varies 2.0× for cerebras.

Measured time to first token, by region

Requested fromProviderModel answeringTTFT p50TTFT p95Samples
South America (São Paulo)cerebrasgpt-oss-120b757 ms1215 ms57
South America (São Paulo)groqllama-3.3-70b-versatile764 ms1126 ms57
Europe (Germany)cerebrasgpt-oss-120b801 ms1398 ms58
Europe (Germany)groqllama-3.3-70b-versatile899 ms1370 ms58
South America (São Paulo)openrouternvidia/nemotron-3-nano-30b-a3b:free1037 ms1343 ms57
South America (São Paulo)googlegemini-flash-lite-latest1143 ms1724 ms57
Asia (Tokyo)groqllama-3.3-70b-versatile1280 ms1624 ms57
Europe (Germany)openrouternvidia/nemotron-3-nano-30b-a3b:free1282 ms2009 ms58
Asia (Tokyo)openrouternvidia/nemotron-3-nano-30b-a3b:free1292 ms1761 ms57
Asia (Tokyo)cerebrasgpt-oss-120b1337 ms1669 ms57
Europe (Germany)googlegemini-flash-lite-latest1351 ms2384 ms58
US (Central)groqllama-3.3-70b-versatile1355 ms1984 ms57
US (Central)cerebrasgpt-oss-120b1485 ms1841 ms57
Asia (Tokyo)googlegemini-flash-lite-latest1488 ms1827 ms57
US (Central)googlegemini-flash-lite-latest1498 ms2140 ms57
US (Central)openrouternvidia/nemotron-3-nano-30b-a3b:free1570 ms1981 ms57

Real streamed completions, last 24 hours, requested from each probe region. The model is named on every row because it is the dominant factor in the number next to it.

What is the difference between time to first token and time to first byte?

Time to first byte (TTFB) is how long it takes the network and the provider’s edge to begin responding at all: DNS resolution, TCP connect, TLS handshake and the first byte back. It does not involve the model. Time to first token (TTFT) is how long a real streamed completion takes to emit its first token: that same network time, plus queueing at the provider, plus the model’s prefill over your prompt.

TTFB is the floor — you pay it on every request no matter which model you call, and no prompt engineering removes it. TTFT is what a human actually waits for before words appear on screen. They are different questions and they have different answers, which is why this site measures both and never averages them together.

How much of time to first token is network latency?

Between 3% and 33% of time to first token is network time — the rest is the provider queuing your request and the model generating its first token. That is why a TTFT number quoted without the model name tells you almost nothing about the API: most of what it measures is the model.

Requested fromProviderTTFT p50TTFB p50Network share of TTFT
South America (São Paulo)cerebras757 ms183 ms24%
South America (São Paulo)groq764 ms235 ms31%
Europe (Germany)cerebras801 ms201 ms25%
Europe (Germany)groq899 ms296 ms33%
South America (São Paulo)openrouter1037 ms58 ms6%
South America (São Paulo)google1143 ms160 ms14%
Asia (Tokyo)groq1280 ms146 ms11%
Europe (Germany)openrouter1282 ms100 ms8%
Asia (Tokyo)openrouter1292 ms56 ms4%
Asia (Tokyo)cerebras1337 ms196 ms15%
Europe (Germany)google1351 ms101 ms7%
US (Central)groq1355 ms107 ms8%
US (Central)cerebras1485 ms68 ms5%
Asia (Tokyo)google1488 ms56 ms4%
US (Central)google1498 ms43 ms3%
US (Central)openrouter1570 ms56 ms4%

Can providers be ranked by time to first token?

Not across different models, and that is the honest answer rather than a hedge. Each provider here is measured on whichever model that account can call — the models are named in the table above — and model size drives first-token time far more than the network or the provider’s hardware does. A ranking built from those numbers would be a ranking of models wearing provider names.

Measured from South America (São Paulo), openrouter has the lowest edge latency (58 ms TTFB) while cerebras has the lowest time to first token (757 ms TTFT, on gpt-oss-120b). Same region, same probes, same schedule — two different winners. “Fastest AI API” is not a question that has one answer until you say which of the two you are paying for.

Which providers have measured TTFT here, and why not all of them?

Inference is measured for 4 of 45 providers (cerebras, google, groq, openrouter), because a streamed completion needs a paid API key for that provider and every probe spends tokens. Edge latency needs no key, so TTFB is published for all 45. We would rather publish 4 measured TTFT figures than estimate 45: an estimate presented as a measurement is the one thing this site does not do.

Method in full, including limits: methodology · Edge latency for every provider: fastest AI API by region · Availability: uptime comparison · Raw numbers under CC BY 4.0: JSON API.