You add streams. Nothing else about the bill moves.
$150
per concurrent stream, whatever it speaks
Most platforms sell a share of one pool with a meter on top, so growth costs twice. Here you add a stream when you need one more thing talking at once, and nothing else moves.
What a stream is
- One stream is one synthesis at a time. Two conversations at once need two streams; a hundred need a hundred.
- Inside a stream nothing is counted: characters, minutes, requests, voices and languages are all unmetered, so the bill is the same on a quiet week and a record one.
- Overflow spills rather than fails, burst streams cover a spike by the day, and automatic failover is wired beside your existing vendor.
The figure, drawn
fig.
The window first audio has to sit inside
Fig
Human handoff · the window to stay inside
~200 ms
one second
The single-stream run that stood beside this window is: first audio byte in 146 ms over the open internet, 116 ms p50 first audio, server side warm. The harness is on the latency benchmark page, run your own shape against your own key.
Past your streams
A live stream is never thinned to make room for a new one. Traffic beyond the streams you hold spills into burst, which charges by the day, the honest answer to a Monday that looks nothing like a Sunday.
How to check us
The single-stream methodology is on the latency sheet, and the harness that produced the figures above is published with them. Point it at your own key, at the concurrency and hours you need, and read your own numbers.
The harness is the public API. Take a key and rerun it with your script in it.
Get a key and rerun this yourself