Time to first token, measured in the open.
A benchmark of typical voice turns runs against the live endpoint around the clock. This page shows everything it has measured. It refreshes every minute.
No results yet
The benchmark has just started, or its results could not be loaded. Check back in a few minutes.
Method
- What is timed. Time to first token runs from the moment a request reaches the endpoint to the first real output (text or a tool call). Network time between you and the endpoint is not included; every API response reports its own figure in the
x-ttft-msheader. - What is sent. Turns a phone agent handles on a real call: an opening line, a mid-call reply, a reply late in a long call, a turn that calls a tool, the reply after a tool result, a reply after an interruption, and an inbound receptionist answering from a long knowledge prompt. Prompts are about 1,000 to 1,500 tokens with tool definitions, and replies are capped at 120 tokens.
- Cold by default. Each turn uses fresh call details and a new session, so its prompt has not been seen before. One scenario repeats a turn inside the same session to show a call already in progress.
- Even spacing. Turns run at a steady interval around the clock, with scenarios taking turns, so every hour and every scenario is sampled alike.
- Nothing dropped. A turn that gets no answer counts as slower than every answer. Percentiles shown as “no answer” mean that share of turns got none.
- One provider alone. Each benchmark turn also records how long individual providers took on the same request. The comparison row uses the median of them for that turn.
The same summary is available as JSON from the API at /benchmark, without a key. API reference.