Gemini 3.8 Flash TTS designs a voice from a prompt, replicates one from 30 seconds, and bills about half the 3.1 rate (2026)

Google published two text-to-speech models on September 23, both in general availability: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. They are audio-only models built on the Gemini 3 Pro base and replace a 3.1 Flash TTS that never left preview. 1 2

Signal	Confirmed detail	What it means for you
What shipped	Two generally available models on the Gemini 3 Pro base: text in up to 8K tokens, audio out. 2	Unlike the 3.1 preview, this is a surface you can build on.
What it costs	Flash TTS bills $0.50 per 1M input and $9.00 per 1M audio tokens; Flash-Lite $0.50 and $6.00. Both double on January 1, 2027. 3	At Google's stated 25 audio tokens a second, that is roughly $0.014 a minute on Flash and $0.009 on Lite, against $0.03 on the 3.1 preview.
What breaks	Unary requests now return WAV instead of headerless PCM, and the text field is read strictly as a verbatim transcript. 4	Pipelines that prepended a WAV header, or wrote stage directions into the script, need a rewrite.

Direction becomes part of the script

What separates the tiers is voice design: Flash TTS invents a vocal identity from a natural-language description and covers 130 languages against Flash-Lite's 101, while Flash-Lite skips it and is positioned for high-volume dubbing and voice agents. Google's launch post puts the library at 2,000+, the DeepMind page says "over a thousand", and the API documents list 30 studio voices by name. 1 4 5

Delivery control splits in two. Sustained direction — "whispering", "out of breath" — belongs in a per-turn speech_metadata.style field; momentary vocal bursts and pauses stay inline as angle-bracket tags in the transcript. Both models share one schema, so changing tier is a single parameter. 4

Two speakers can be staged in a single request with mode: "conversational", but only using prebuilt voices. Mixing custom or replicated voices in one dialogue means synthesizing each turn separately and concatenating. 4

Voice replication rebuilds a speaker from a 30-second sample, behind a consent recording from the voice's owner. Stored voices are capped at 200 per project with a one-year expiry, and stateless voice keys last seven days. Replication through AI Studio is unavailable in Illinois, Texas, the EEA, the UK, Switzerland and India, and every generated clip carries a SynthID watermark. 1 4

Horizontal bar chart of the Hume AI Overall Quality Index for eight text-to-speech models, with Gemini 3.8 Flash TTS at 0.920 and Flash-Lite at 0.914 topping the field — Google's published comparison on Hume AI's Overall Quality Index, 23 September 2026: Gemini 3.8 Flash TTS and Flash-Lite take the top two places among the eight models Google measured, ahead of Cartesia Sonic 3.6 and both ElevenLabs v3 variants. Compiled by this article from Google's published table; the figures are Google's. 5

Where it stands, on Google's numbers

On Voice Arena's blind human Elo for US English, Flash-Lite leads at 1087 and Flash sits third at 1061, behind Cartesia's 1068. The same Google tables show where the new models lose: ElevenLabs v3 rates closer to human turn-to-turn variation, 5.00 against Flash's 4.58, and leads the single-tag voice-qualities measure 76.6 to 74.6. 5 6

All of it is Google's or its partners'; no independent measurement of the released models has appeared yet. 5

Why it matters

Persona, scene and per-line direction become text you can generate, version and reuse, which is what lets a script-to-audio pipeline live inside one API family. The price is the reason to move now, with two caveats: both rates double on January 1, and the voice cap is per project, not per key. 1 3

Flash TTS is rolling out in the Gemini API, Google AI Studio and Gemini Notebook; Flash-Lite in the API, AI Studio and Google Vids. Gemini Enterprise is "coming soon". 1