Sonic 3.6, 3.5 and 3 cost the same on Sume TTS Router: pick by ear

All five Sume TTS Router ids share one per-character price, so model choice is a listening test, not a budget one. Here is how to run it for a few cents.

4 min readSume
All posts

On Sume's TTS Router, sonic-3.6, sonic-3.5 and sonic-3 carry the same per-character price, so you choose between them by listening, not by budget. Render the same 600-character passage on each and the whole comparison costs about 9 cents. TTS 1.0, the managed route, always uses sonic-3.6 and has no model field.

What the catalog lists

The router seeds five ids, all Cartesia Sonic. Each has the same list price per character and the same 1.25 margin. sonic-latest is an alias that resolves to sonic-3.6, the provider's current stable release, never to the preview. sonic-preview is a provider beta channel: output and availability can change without notice, and it rejects pro voice clones with voice_model_mismatch.

Sume rates and limits, read 2026-10-06 from the repo catalog and schemas
Model idWhat it isPrice per 1,000 charactersCaveat
sonic-3.6Current stable Sonic$0.0475Used by TTS 1.0
sonic-3.5Previous Sonic$0.0475Pin it to keep a series sounding the same
sonic-3Older Sonic$0.0475Same voices, older model
sonic-latestAlias for sonic-3.6$0.0475Moves when the provider's stable release moves
sonic-previewBeta channel$0.0475Can change without notice; no pro clones

A blind test that costs cents

Pick one passage of about 600 characters that has your hard cases: numbers, a brand name, a question, a long sentence. At the rate above, a 600-character job is 2.85 cents, which bills as 3 cents. Three models is 9 cents. Run each with the same voice id and the same language, export the files with the same settings, rename them A, B and C, and have two people rank them without knowing which is which.

  • Same voice id, same language, same speed and volume for every take.
  • Hide the model names before anyone listens.
  • Include a numbers sentence and a name sentence, since those fail first.

Reading the results

Score each take on three things, not an overall impression: did it say every number and name correctly, did the pacing match the script's punctuation, and would you be happy hearing it for ten minutes. A model that wins on a sixty-second sample can still tire a listener over a longer piece, so if the script is long, test a longer passage for the finalist.

Write down the winner's model id, the voice id, the language and the date. If the provider moves its stable release later, that note is how you reproduce today's sound.

Pin or float?

If a series has to sound identical across episodes, name a specific model id such as sonic-3.5 rather than sonic-latest, and record it with the script. If you always want the newest stable model, sonic-latest does that for you but can change timbre when the alias moves. Avoid sonic-preview for anything you are shipping, since the provider can change it without notice.

Sources

Related posts

More in Models

All Models posts

Written by Sume