Gemini 3.8 Flash TTS context caching vs Sume transcripts

Gemini 3.8 Flash TTS lists a context-caching rate of $0.125 per million tokens through 2026-12-31, rising to $0.25 on 2027-01-01 (read 2026-10-07). The cache applies to the text input side, not to the audio you get back, so it only helps if you resend a large shared prefix. Sume TTS jobs take a transcript of 1 to 20,000 characters each time and have no cache setting.

What the Gemini cache price covers

Gemini bills TTS on two sides: text input tokens and audio output tokens. Context caching discounts the repeated input. At $0.50 per million input tokens against $0.125 cached, a long reused prefix costs a quarter as much to resend. But audio output at $9.00 per million tokens through 2026 is the larger line for narration, and the cache does not touch it.

When a cache would ever pay off

Caching pays off when many requests share a long identical front section, for example a style brief repeated before each line. For ordinary voiceover the input is the script itself, and each line differs. A short ad line has so few input tokens that a cache saves fractions of a cent next to the audio output cost.

So before you design around a cache, estimate your input-to-output ratio. Google's page prices audio output at $9.00 per million tokens against $0.50 for text input, so measure your own ratio on a real script.

What Sume does instead

A Sume TTS 1.0 job is character-metered. Each request carries its own transcript (up to 20,000 characters) or a transcript_source reference to an accepted script, never both. There is no prefix to cache, and no shared-context field in the request schema.

What you can reuse is the result. Send an Idempotency-Key and an exact retry returns the same job instead of making a second paid one. Reusing a key for a different payload fails with 409 idempotency_conflict.

For per-character list rates on the Sonic models, read the live catalog at GET /v1/tts-router/models rather than a number in a blog post. The router docs state the catalog rate is the provider list price times 1.25.

• Gemini: budget the audio output line first, the cache second.
• Sume: budget characters per job, and use idempotency keys for retries.
• Keep the artifact URL from the first Sume result instead of regenerating the same line.

Decision rule

If you run thousands of short lines with a long shared instruction block, price Gemini with the cache and check the 2027 doubling. If your lines are self-contained scripts, caching is a rounding error on either side and the choice rests on voice, language and where the audio needs to land. Sume's docs do not describe Gemini as a routable engine, so the comparison is between separate vendor accounts.