Google's September 23 launch brings two stable speech models to the Gemini API and AI Studio: Flash for creative work and Flash-Lite for volume. Gemini Enterprise API access is coming later.
Both cost less per generated minute than 3.1 Flash TTS Preview. Their Standard rates double on January 1, 2027. Migration also changes how applications send instructions and save audio.
The big change
- What changed: The new tiers separate transcript text from performance directions and support reusable voices.
- Why it matters: A voice pipeline has two migration costs to consider: engineering changes now and the scheduled increase in generation charges.
- What to watch: For dubbing, audiobooks and voice agents, compare intelligibility and speaker consistency on real production scripts across languages. Those results determine whether a cheaper generated minute also reduces the cost of usable audio.
Output cost per minute
Google's pricing table specifies 25 audio tokens per second: 1,500 per minute. The calculation is output rate × 1,500 / 1,000,000; text-input charges are additional. All figures below are US dollars at Standard rates.
Model | Listed languages | Input / output per 1M tokens through Dec. 31, 2026 | Output/minute now | Output/minute from Jan. 1, 2027 |
|---|---|---|---|---|
Gemini 3.8 Flash TTS | 130 | $0.50 / $9 | $0.0135 | $0.027 |
Gemini 3.8 Flash-Lite TTS | 101 | $0.50 / $6 | $0.009 | $0.018 |
Gemini 3.1 Flash TTS Preview | — | $1 / $20 | $0.03 | No announced change |
For the 3.8 models, Batch and Flex cost half Standard; Priority costs 1.8 times Standard. Google's model index recommends replacing the legacy 3.1 preview with either new tier.
Prompts and audio files need changes
The Flash migration guide says transcript text is spoken verbatim. Put sustained directions and speaker labels in speech_metadata; an old instruction such as “Say cheerfully:” can otherwise be read aloud. Brief events such as <laugh> remain inline.
Every multi-speaker turn must name a speaker matching the configuration. Unary responses now contain WAV headers, replacing the old default of raw PCM. Remove code that adds a WAV header, or explicitly request a raw format.
Flash-Lite's documentation confirms a shared schema and limits of 8,192 input tokens and 16,384 output tokens for both tiers. Applications can therefore compare them through the same integration.
Voice replication and evaluation
Google describes more than 2,000 ready-made voices, voice design and replication from a 30-second sample of a voice the user has rights to use. Replication requires a matching verbal-consent recording. Google says generated clips carry SynthID, and replicated voices have C2PA credentials.
AI Studio voice replication excludes Illinois, Texas, the European Economic Area, the UK, Switzerland and India.
Google reports Flash scores of 71.4 overall and 60.8 for accent modeling on Hume's Voice Design Benchmark. It places Flash and Lite first and second on Hume's Overall Quality Index.
Artificial Analysis separately listed Flash at 1260.15 Arena Elo, 27th among 95 models, on September 23. This measures crowd preference; the reviewed listing does not establish performance by language or over long recordings. No equivalent independent Lite result was used here.
For a comparison with another hosted speech lineup, see our Qwen Audio 3.1 API and pricing analysis.
