Gemini 3.8 Flash TTS: design a voice, prove consent

Google's new text-to-speech models let you write a voice into existence — and they will not clone a real one without a recorded consent statement. On September 23, Google shipped Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS to the Gemini API and AI Studio. The interesting part of this release is not the audio quality. It is that Gemini 3.8 Flash TTS ships provenance and consent as product features, not policy pages — and that the introductory pricing has a hard expiry date.

What actually happened

Two models, split by job. Flash TTS targets voice design, character acting and long-form audio, with line-by-line control over delivery, pacing, dialect shifts and scripted vocal bursts. Flash-Lite TTS targets high-volume work — dubbing, bulk audio, voice agents — at a lower output rate.

Voice design is the headline capability: describe a role, accent and set of vocal characteristics in plain language, across 100+ languages and regional varieties, and get back a reusable voice ID you paste into your code. No more picking from a fixed preset list.

Voice replication is the part with rules attached. Google recreates a consistent vocal profile from a 30-second sample, but the flow requires a verbal consent recording from the voice owner that matches the reference speaker before the voice can be created. Every clip the Gemini audio models generate carries SynthID watermarking, and replication adds C2PA credentials.

Now the meter. Google's pricing page lists text input at $0.50 per million tokens for both models through December 31, 2026, with audio output at $9.00 per million tokens for Flash TTS and $6.00 for Flash-Lite TTS. On January 1, 2027, every one of those figures doubles: $1.00 input, $18.00 and $12.00 output. There is a free tier for both. Enterprise availability through the Gemini Enterprise API is listed as coming soon.

Why voice design and consent verification matter for your business

Built-in consent is a liability control you were otherwise going to build yourself. If you have ever wanted your founder's or receptionist's voice on a phone agent, the hard part was never the audio. It was the paperwork and the proof. A platform-enforced consent recording plus C2PA credentials on the output means the provenance trail exists without you inventing a process for it. Keep your own record too — but the default is now on your side.

SynthID changes what your voice content is. Every generated clip is watermarked and detectable. If your plan involves synthetic audio that a customer might reasonably assume is a human employee, decide now how you disclose it. Detectability is not a future risk. It shipped.

Price the workload in minutes, not tokens. At Google's audio token rate, the output cost lands under two cents per minute of speech today, and roughly double that from January. For an IVR that plays the same twelve prompts a thousand times a day, cache the audio and the model cost is a rounding error. For a live agent generating fresh speech every call, the January rate is the number your unit economics have to survive.

Pin the model in config. Two TTS models with different rates, plus a price change on a known date, is exactly the situation where a hardcoded model string becomes a January surprise. Name the model in configuration, log audio output tokens per workflow, and run the math at the 2027 rate before you call a voice product economical.

Sources: Google, Gemini API pricing.