On a meter · 82 characters
“Rebooked on the 6:40. Gate B12. Anything else?”
Correct, complete, and forty percent cheaper to say. It answers the question and none of the feeling underneath it.
Why Gandr
A meter is a tax on personality.
Per-character pricing looks like a billing decision. It is not, it is a product decision you did not get to make, applied to every sentence your product speaks.
The same call, two pricing models
This is what the meter deletes.
A traveller has missed a connection. Same customer, same situation, same agent, one of them is being billed by the character.
On a stream · the same price
“Okay, take a breath, you are not going to miss it. I have already moved you to the six-forty, window seat, and your bag is checked through. Gate B12. Go straight to security and I will stay on with you until you are through.”
Same information. The difference is the part a meter charges you extra for: the reassurance, the specifics, the promise to stay on the line.
Both agents solved the problem. Only one of them sounded like it cared, and on a meter that one costs more every single time it runs.
You cannot meter a voice into sounding human.
Our founder, on why we are working on this:
The biggest mistake people make with AI speech is thinking it's about making robots sound pretty. It's not. Speech is the highest-bandwidth protocol humans have for transferring complex intent. If you want a frontier model to actually run the physical world, it can't just process code. It has to master the underlying frequency of interaction. We aren't building a voice tool. We are building the conscious protocol that lets autonomous systems finally understand, adapt, and operate in our reality.
What changes
Buy the stream, not the sentence.
A stream is one thing talking at a time, and nothing it says is counted. On a metered plan you buy a seat and a meter runs above it; here the stream is the whole bill. Need two things talking at once, buy a second stream. Your agent can say the extra sentence, and nothing happens to the invoice.
- 01
Write the warm version of the prompt. It costs the same as the terse one.
- 02
Let calls run to their natural end. The long ones are where the product learns.
- 03
Forecast voice as a fixed line, not as a function of how much your customers talk.
The answer, itemized
first audio 116 ms p50 first audio, server side warm
The breath. First audio inside a breath →
The calls above run on the same streaming surface the API sells.
The stream. One stream, nothing metered →
Every line here rides one stream, billed flat, what you run through it is not counted.
The clone. A voice from ten seconds →
Ten seconds of reference audio casts a new voice on any stream.
Take one stream and hear the difference.
Sign in, pay, and a key is on your screen, one concurrent stream on the production API, with nothing on it counted. Write the warm version of your prompt and listen to what you have been paying to delete.
Get a keyAt 500 lines and more there is no published number. Tell us your peak and we price it. Book a 30-minute call →
Calendar broken? [email protected]