Every turn is a race. Your caller hears the winner.
Any one model provider has slow moments: a queue, a cold cache, a bad minute. A coding tool absorbs that. A caller hears dead air and talks over your agent.
For every turn of DeepSeek V4.1 Flash, FasterTokens queries multiple inference providers simultaneously and optimizes for time to first token. It is one OpenAI-compatible endpoint, and its latency is measured around the clock and published.
Each dot is one conversational turn, placed by how long after the request arrived the first token left.
Against one provider on its own.
On every benchmark turn we also record how long a single provider took to start answering the same request. Alone, a provider has a slow tail. The bars show the 95th percentile: 19 of every 20 turns were at least this fast.
31 measured turns, November 12, 2026 to November 12, 2026. “One provider alone” is the median of the providers timed on each turn. Every turn is a cold request unless its scenario says otherwise. Method, scenarios and results by hour.
Fast models exist. A fast turn, every turn, takes more than one provider.
Many providers, one endpoint
For every turn we query multiple inference providers at the same moment and stream you the fastest start. When one of them has a bad minute, your caller does not hear it.
Tuned for the first token
Routing is optimized continuously for time to first token, using live measurements of the providers. You get the benefit without tracking who is fast today.
Built for a spoken turn
Reasoning is off, completions are capped at the length of something a person would say, and a turn that gets no first token within two seconds fails fast with a clear error so your agent’s fallback can take over.
Measured in the open
A benchmark of typical voice turns runs against the live endpoint all day, every day. Its results are public, and every response tells you its own time to first token, so you can check ours against yours.
A written guarantee Coming
We are measuring a full week, every hour of every day, before putting a number in writing. The benchmark is public while it runs, so you can watch the same data we will base the guarantee on.
How the number is measured.
Timed at our endpoint
The clock runs from the moment a request reaches our endpoint to the moment the first real output leaves it. Every response carries that time in a header, so you can check it against your own.
Typical voice turns
Call openings, mid-call replies, long calls, tool calls, replies after a tool result, interruptions and inbound reception. Each turn uses fresh call details, so nothing is answered from a warm cache unless the scenario says so.
All day, all week
Turns are spaced evenly around the clock, so busy hours and quiet hours count the same. A turn nobody answers is counted as the slowest, not dropped.
Fits the stack your agent already runs on.
For teams building agents
FasterTokens is a single model endpoint that speaks the OpenAI chat completions format, with streaming and tool calls. Point your framework’s OpenAI-compatible model setting at it and keep your speech recognition, voices and telephony as they are.
For voice platforms
Your customers blame you for dead air, even when the model caused it. Offer FasterTokens as a model option and show them the public benchmark behind it. Write to us about volume and limits.
Change one URL.
Create a key, set the base URL, and send the turns you already send. Add a session ID per call so its turns are served consistently. The preview is free, with a monthly allowance per account.
Read the quickstartcurl https://api.fastertokens.com/v1/chat/completions \
-H "Authorization: Bearer $FASTERTOKENS_API_KEY" \
-H "Content-Type: application/json" \
-H "X-Session-Id: call-8841" \
-d '{
"model": "deepseek-v4.1-flash",
"stream": true,
"messages": [
{"role": "system", "content": "You are a phone assistant. Reply in one short sentence."},
{"role": "user", "content": "Hello?"}
]
}'Stop waiting on one provider.
Get a key in a minute and put it behind your next call. No card needed during the preview.
Get an API key