Gemini TTS models leave preview, with audio output priced at $6-9 per million tokens

Google has made Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS generally available in the Gemini API, according to the Gemini API changelog entry dated September 22, 2026. The Flash-Lite model replaces the 3.1 Flash TTS preview.

The entry appeared on the changelog page, which now redirects to aistudio.google.com/docs/changelog. We found no separate blog post from Google on it.

Two models, one API schema

The changelog describes gemini-3.8-flash-tts as the flagship creative model. Google lists studio-grade voice fidelity, nuanced acting, regional dialects and stable long multi-turn speech as its strengths.

gemini-3.8-flash-lite-tts is the fast, cost-efficient option, aimed at high-throughput production and real-time voice agent cascades. Google’s entry says it is “built to replace gemini-3.1-flash-tts-preview“.

Google’s model page for Flash-Lite adds these specifics:

• Text goes in and audio comes out, with limits of 8,192 input tokens and 16,384 output tokens per request.
• It supports over 100 languages (101 in the comparison table) with automatic language detection. The sister Flash TTS model supports 130.
• It works with the Batch API, Flex inference and Priority inference. It does not support function calling, the Live API, structured outputs, thinking or caching.
• Unary requests now return WAV with a RIFF header by default. Earlier models returned raw PCM, so existing code may need a change. Headerless formats such as audio/l16 remain available.
• Both models share one schema, so switching between them means changing only the model parameter.

Custom voices come with a consent check

The same entry introduces a Voices endpoint, /v1beta/voices, and an Extended Voice Library that Google says holds 150+ prebuilt and custom voices. Developers can use voice design to create a persistent persona from a text prompt. They can also use voice replication, which Google says includes consent verification.

Google’s changelog does not say how that verification works or what it requires from the person whose voice is copied. For a feature that clones voices, that is the main open question. The page also gives no detail on watermarking or misuse policy.

Audio output costs less than half the preview rate

Neither Google page we read states a price. LiteLLM’s day-0 support post, published September 23, 2026, lists launch pricing per million tokens. Flash TTS costs $0.50 for text input and $9.00 for audio output. Flash-Lite TTS costs $0.50 and $6.00. The 3.1 Flash TTS preview it lists for reference costs $1.00 and $20.00.

On those figures, audio output is 55% cheaper for Flash TTS and 70% cheaper for Flash-Lite than the preview. Input costs half as much. LiteLLM also says Batch and Flex run at half these rates and Priority at 1.8 times. It says launch pricing runs through December 31, 2026, and that Google’s standard pricing applies from January 1, 2027.

These numbers come from a third-party gateway, not from Google, so developers should check them against Google’s pricing page. The January change is the one to watch: the post does not say what the standard rates will be.

The cheaper tier matters because a voice agent produces a lot of audio. Output tokens are the larger line item in a cascade that goes from speech recognition to a language model to speech. A cut from $20 to $6 per million output tokens changes the cost of long calls and dubbing. LiteLLM also notes the 8,192-token input cap, so long scripts such as audiobook chapters must be split across calls.

Google’s own claims about fidelity and acting are unverified. The sources include no benchmarks or latency figures against the preview or rival voice services. The comparison with the preview rests on third-party pricing, and the entry sets out no regional availability. The 3.8 TTS models are not covered by the Live API, so real-time conversational audio would still go through a text-to-speech step. Google’s own language figures also differ between the sources we read, from “over 100” to 101 for Flash-Lite.