What a call costs
Every call carries its own cost breakdown: a flat platform fee, then one line per provider that did work. A real call, taken apart.
5 min readUpdated 26 August 2026
Nothing here is estimated. Open Call Log, click a call, then View full detail. Under Call Analysis there is a Cost Breakdown, and it shows every component that was charged.
The platform fee is flat
The first line is Platform Fee, described on the page as "fixed infrastructure charges covering compute, routing, and system overheads". It is $0.004, and it is charged once per call.
Once per call, not per minute. A call that runs six and a half minutes carries the same $0.004 as one that lasts thirty seconds. On a real call in the log lasting 6 minutes 35 seconds, the platform fee reads $0.004; per minute it would have been around $0.026.
Tip
This is why lots of very short calls are cheaper than the fee makes them look, and why the fee is almost irrelevant on a long one.
Then one line per provider
After the fee, the breakdown lists each piece of work separately. Each line names the provider, the exact model, how many units were consumed, what kind of unit those were, and the cost.
On a realtime voice to voice call that means lines such as token output audio and token output text, both against the same model, because a realtime model handles speech in and speech out itself. On a standard pipeline you see transcription and speech synthesis as their own lines instead.
| Line | Value |
|---|---|
| Platform Fee | $0.004, flat |
| Token output audio | 4,398 units, $0.131940 |
| Total Units | 210,238 |
| Total Cost | $0.450966 |
Reading it when a call looks expensive
The breakdown tells you which part to change, which is the entire reason it is itemised rather than shown as one number.
- One provider line dominating means the model choice is doing it, not the call length.
- Everything proportional but large means the call simply ran long, which is usually a prompt that does not close the conversation out.
- A transcription line much larger than expected usually means callers on noisy lines.
Why per minute is the wrong unit
Because there is no single rate. Units are tokens, characters and seconds, priced by each provider for each model. A minute is just how long those units took to accumulate.
This is also why the pricing calculator on the website lets you swap each component. Change the model and the per minute figure moves, because the components are what is actually being paid for.
Common questions
- Is silence charged?
- The models are only charged for the work they do, so a quiet stretch costs less than a busy one. The flat platform fee is unaffected either way, because it is per call.
- Why do two calls of the same length cost different amounts?
- Different models, or a different amount of talking inside the same duration. Open both breakdowns and compare line by line.
- Does a failed call cost anything?
- It costs whatever work actually happened before it failed. Open the call and look. A call that never connected has very little against it.
Was this helpful?