Why it won the phone job
A phone network carries 8 kHz mu-law audio. Most speech APIs return MP3 or 24 kHz PCM, and you convert. Cartesia lets you ask for pcm_mulaw at 8000 Hz in the request. The bytes that come back go straight into the Twilio stream with no transcoding step on my server. On a live call that removes one whole stage where delay and bugs can hide.
It also streams. My code sends text to Cartesia as the language model produces it and forwards audio to the caller as it arrives. The caller starts hearing the answer while the end of the sentence is still being generated.
Two small rules I ended up with:
- Skip any text fragment shorter than 3 characters. When the model streams, the splitter sometimes emits a stray comma or a single letter, and there is nothing there to voice.
- Generate the greeting and the filler phrases once, keep the audio in memory, and replay it. Those lines never change, and they are the ones that have to play instantly.
One voice, used everywhere
I standardized on a single stock voice at speed 1.0. I tried 0.85 for a walkthrough video and it dragged, so 1.0 stayed. The same voice read the receptionist's calls, the product demo on its landing page, and outreach videos I made for prospects.
One of those videos runs 64.7 seconds and its script is 1,045 characters. Cartesia bills one credit per character, so the whole narration cost 1,045 credits. On the $49 plan that is about 4 cents.
For narration I do not stream. A script calls the tts/bytes endpoint once per line, saves each clip, measures its length, and builds a timeline from the real durations. The video is cut to the voice, never the reverse.
The "Model sunsetted" error
I had every script pinned to the sonic-2 model. Then I voiced a demo in Italian and the API returned a 400 with the message "Model sunsetted." English requests on the same model still worked. Cartesia had retired sonic-2 for other languages first.
The fix was sonic-3 and a native Italian voice. The lesson is about pinning. Pin the model so your voice does not change under you, and put the model name in one config value so the day it retires costs you one edit.
Credits and what a call minute costs
| Plan | Monthly price | Credits | TTS requests at once |
|---|---|---|---|
| Free | $0 | 20,000 | 2 |
| Pro | $5 | 100,000 | 3 |
| Startup | $49 | 1,250,000 | 5 |
| Scale | $299 | 8,000,000 | 15 |
Sonic text-to-speech uses one credit per character. Cartesia's own page says a minute of audio is about 750 to 800 credits.
On the Startup plan a credit costs $0.0000392. A minute of nonstop speech is about 3 cents. A phone agent does not talk nonstop. On my line the agent spoke roughly half the time, about 450 characters per call minute, which works out to 1.8 cents a minute for voice.
Watch the last column. Five requests at once on Startup means five callers can be hearing speech at the same moment. A sixth request is over the limit. For a single business line that is plenty. For a call center it decides your plan before price does.
Cartesia now sells a full voice-agent product at $0.03 a minute plus telephony and model costs. I have not built on it.
Why my narration left
In October 2026 my Cartesia credits ran out. I tested Kokoro, a free model that runs on my Mac, on the same scripts. For narration it was good enough that I made it the default on October 28. The Kokoro page covers what I had to build around it.
Cartesia still does the two jobs Kokoro can't do for me: live calls, where streaming and phone encoding matter, and videos that have to match the voice of earlier ones.