Sonic 3.6 vs 3.5 vs 3 on Sume TTS: same price, so pin the id

Sonic 3.6, Sonic 3.5 and Sonic 3 cost the same on Sume: 38 micro-dollars per character at list, 47.5 billable, with a 20,000-character cap per request. Because price does not separate them, pin an explicit id (for example sonic-3.6) instead of sonic-latest if you need the next episode to sound like the last one.

What the catalog lists

The Sume TTS router publishes its catalog at GET /v1/tts-router/models. Every entry is pass-through text-to-speech and all five ids below share the same per-character rate and request cap. The only differences are in the constraints column.

Why not just use the alias

sonic-latest is documented in the catalog as an alias for sonic-3.6, the provider's current stable Sonic release. It is never sonic-preview and not a dated snapshot. When the provider ships the next stable release, an alias can move; a pinned id does not. For a series where voice continuity matters (a weekly show, a course with 40 lessons), the pinned id is the safer default.

A completed text-to-speech job records model_id, the voice, the language, the output format and the synthesis settings. Store those values with each episode so a later retake can request the same engine.

A pinning checklist

Pin the id in your request body rather than in a config alias, and write it down next to the episode. Store model_id, voice, language, output format, generation_config and speed from the completed job. If a null appears for a setting, the request did not send it, so a retake that sends nothing will behave the same way.

Test a new Sonic release on a short, representative script (a 600-character paragraph costs 3 cents: 600 x 47.5 = 28,500 micro-dollars, 2.85 cents, rounded up) before you change the pinned id for the whole series.

When preview is the wrong choice

sonic-preview is the provider's beta channel; output and availability can change without notice. The catalog also says it is not compatible with pro voice clones: those requests fail with voice_model_mismatch. Use it to test, not to ship.

At the shared rate, none of this changes the bill. A 1,000-character job costs 47,500 micro-dollars on any of the five ids, which rounds up to 5 cents.

What to check before the next release

When a new Sonic id appears in the catalog, the first thing to compare is the per-character rate and the request cap, both visible in GET /v1/tts-router/models. If the rate is unchanged, you can A/B the new id on a representative script for pennies and decide on sound alone.

If you use a cloned voice, also read the constraints column: preview is documented as incompatible with pro voice clones, and a request that mixes them fails with voice_model_mismatch.