Cartesia hosted LLMs free until Oct 1: price just the voice on Sume

Cartesia's changelog says its Managed Agents come with hosted LLMs that were free until October 1, 2026, and that Line SDK agent hosting ends on December 1, 2026. If you built a talking agent there, the bill just changed shape. If all you used was the voice, you can price that part on its own: Sume's TTS router serves Cartesia Sonic models at $0.0475 per 1,000 characters, and it is a job API, not an agent host.

What the changelog says

The August 2026 entries list Sonic 3.6 (released Aug 27), the API version 2026-08-14 with new locale, accent and normalization fields, multilingual voice clones, Managed Agents with no backend required, and the end of Line SDK hosting on Dec 1. The free-use date for hosted LLMs is the one that already passed. Sume does not host agents or LLM turns inside the TTS router, so this post covers the voice line only.

Pricing only the voice

Take a support agent that speaks 400 replies a day averaging 250 characters. That is 100,000 characters a day. At $0.0475 per 1,000 characters it is $4.75 a day before rounding per reply. Each 250-character reply bills 250 x 0.00475 = 1.1875 cents, which rounds up to 2 cents, so the real figure is 400 x 2 = $8.00 a day. Per-job cent rounding matters for short replies, and the TTS router is an asynchronous job API without streaming, so this is a fit for prepared audio, not live conversation.

replies, chars = 400, 250
cents = max(1, -(-chars * 475 // 100_000))
print(cents, "cents per reply;", replies * cents / 100, "dollars per day")

A fair way to decide

If the product is a live voice agent, you need streaming and turn-taking, and Sume's router lists streaming TTS as a non-goal, so keep the agent platform. If the product is prepared audio, such as narrated clips or notifications generated in batches, price it per character, and the per-reply rounding above tells you where batching longer scripts saves money.

Either way, write down the date the free period ended next to your cost model, so the sudden change in a bill has an explanation.

What to write in your cost model

Add one line per component: voice per character, language-model turns per token, platform fee per minute if any. Mark each with the page it came from and the date you read it. The Cartesia entry shows why: a component that was free until a date silently becomes a cost line afterwards, and a model that never recorded the date cannot explain the jump.