Google made Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS generally available on the Gemini API on September 22, 2026, and announced them on its blog on September 23. Both models turn a text script into audio. The new parts are voice design from a text prompt, voice replication from a 30-second sample with a recorded consent statement, and a promotional price that doubles on January 1, 2027.
The price cliff is the decision. At the promotional rate, an hour of Flash TTS audio costs $0.81. From January 1 it costs $1.62. Anyone budgeting a voice product on today's rate is budgeting at half the price they will pay in fourteen weeks.
Every figure here comes from Google's pricing page, its text-to-speech guide, the API changelog and the September 23 announcement. The benchmark positions are Google's own claims and have not been independently replicated. No latency figure has been published.
- 01Promotional pricing runs to December 31, 2026 and doubles on January 1, 2027.Flash TTS audio output is $9 per million tokens now and $18 from January 1. Flash-Lite is $6 now and $12 from January 1. Text input doubles from $0.50 to $1.00 on both.
- 02Voice cloning needs a 30-second sample plus a spoken consent recording, and is off in six places.Google's footnote excludes Illinois, Texas, the EEA, the UK, Switzerland and India for voice replication through AI Studio. Up to 200 stored voices per project.
- 03Google publishes three voice counts, and they measure different things.30 prebuilt studio voices in the guide, 150+ prebuilt and custom voices in the changelog, and 2,000+ production-ready voices in the announcement.
- 04Prototype now at the promotional price. Budget at the January 1 price.The per-minute cost is arithmetic from Google's per-10-second equivalents. Flash TTS after January 1 is only 10% cheaper per token than the 3.1 preview it replaces.
01 — The releaseWhat shipped #
- Model IDsText in, audio out. Default output is WAV.
- gemini-3.8-flash-tts · gemini-3.8-flash-lite-tts
- Input / output tokensPer Google's model pages. Audio is billed at 25 tokens per second.
- 8,192 / 16,384
- LanguagesPer the text-to-speech guide. The announcement says over 100 for the pair.
- 130+ Flash · 100+ Flash-Lite
- Ways to get a voicePrebuilt, Extended Voice Library, voice design from a prompt, voice replication from a sample.
- 4
- Multi-speaker in one requestPrebuilt voices only. Custom voices are synthesised one turn at a time and joined.
- Up to 2 speakers
- Where it runsGemini Enterprise, the renamed Vertex platform, is listed as coming soon.
- Gemini API · AI Studio
Google positions the two models differently. The changelog describes Flash-Lite TTS as the model built to replace the 3.1 Flash TTS preview for high-throughput production and real-time voice agent cascades. Flash TTS is pitched at studio-grade voice work, acting, regional dialects and long-form audio. Neither model supports function calling, the Live API, thinking, structured outputs or grounding, so a voice agent still needs a separate text model in front of it. Our post on choosing between Gemini 3.8 Live and a cascade covers that split.
Both models are stable releases. Google's deprecations page names them as the replacement for the 3.1 Flash TTS preview and for the two 2.5 TTS previews from May 2025, and lists no shutdown date for any of the three. Every clip carries a SynthID watermark.
02 — The invoiceThe price now and on January 1 #
Google's pricing page lists every paid rate twice: one figure through December 31, 2026, and one starting January 1, 2027. The second figure is exactly double the first on every line. Prices are per million tokens, input then output. Google bills audio at 25 tokens per second, so a minute of audio is 1,500 output tokens.
| Source: Google Gemini API pricing page. Per million tokens, input / output. The per-minute column is our arithmetic from Google's per-10-second equivalents, promotional then January 1 rate. | |||
|---|---|---|---|
| Model and tier | To Dec 31, 2026 | From Jan 1, 2027 | Per audio minute |
| --- | --- | --- | --- |
| Flash TTS · Standard | $0.50 / $9.00 | $1.00 / $18.00 | $0.0135 → $0.027 |
| Flash TTS · Batch or Flex | $0.25 / $4.50 | $0.50 / $9.00 | $0.00675 → $0.0135 |
| Flash TTS · Priority | $0.90 / $16.20 | $1.80 / $32.40 | $0.0243 → $0.0486 |
| Flash-Lite TTS · Standard | $0.50 / $6.00 | $1.00 / $12.00 | $0.009 → $0.018 |
| Flash-Lite TTS · Batch or Flex | $0.25 / $3.00 | $0.50 / $6.00 | $0.0045 → $0.009 |
| Flash-Lite TTS · Priority | $0.90 / $10.80 | $1.80 / $21.60 | $0.0162 → $0.0324 |
| 3.1 Flash TTS Preview · Standard | $1.00 / $20.00 | no change listed | $0.03 |
Three things follow from the table. First, the promotional Flash TTS rate is 55% below the 3.1 preview per audio minute. After January 1 the gap shrinks to 10%, because $18 against $20 is a small saving. Second, Batch and Flex halve every input and output rate and Priority multiplies it by 1.8, in both periods. Third, context caching is also promotional, $0.125 per million now and $0.25 from January 1, which matters for long scripts that reuse a system instruction.
A worked example, illustrative only. A dubbing pipeline producing 1,000 hours of audio a month on Flash TTS Standard pays about $810 for output today and about $1,620 from January. The same volume on Flash-Lite Batch pays about $270 now and $540 later. Input costs are small by comparison, since a minute of speech is a few hundred text tokens. The choice of tier moves the bill more than the choice of month, but both double.
03 — The rulesVoice cloning: consent and geography #
Google calls the feature voice replication. According to the announcement, the model can recreate a vocal profile from a 30-second sample, and the request must include a verbal consent recording from the voice owner. The guide adds that stored voices have a one-year life and a cap of 200 per project, and that a stateless voice key expires after seven days.
The announcement's footnote states that voice replication through AI Studio is not available in Illinois, Texas, the EEA, the UK, Switzerland and India. Google gives no reason in the post. If your users or your voice talent sit in one of them, plan on voice design from a prompt instead, which has no such restriction listed.
Two practical limits sit under the headline. Multi-speaker generation in a single request handles two speakers and only prebuilt voices. A cloned or designed voice has to be synthesised one turn at a time and joined afterwards, which puts the pacing of a two-voice dialogue back on your pipeline. And voice remixing, the ability to shift the timbre or accent of a library voice, is listed as coming soon rather than shipped.
04 — The countWhat the voice count actually is #
Google's own material gives three different voice counts for the same launch. They are not contradictory once you see what each one measures, but a buyer comparing vendors will meet all three.
Prebuilt studio voices
The named voices, Zephyr, Puck, Kore and the rest, that work in multi-speaker requests and need no setup. This is the number OpenRouter's catalog repeats.
Prebuilt and custom voices
What the new Voices endpoint returns when you list the Extended Voice Library. The guide describes the library as hundreds of additional voices across languages, accents and personas.
Production-ready voices
The marketing figure. It does not appear in the guide, the changelog or the model pages, so treat it as the vendor's own description of its library rather than a count you can list.
The same care applies to the benchmark claims. The announcement says Flash TTS took the top overall position on Hume AI's Voice Design Benchmark with a score of 71.4, that the two models hold the first two places on Hume's Overall Quality Index, and that both sit in top positions on Voice Arena's blind preference tests in Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi. No Elo figures or confidence intervals are given. Our guide to reading vendor voice benchmarks explains why a vendor's position on a test it chose is a claim, not a ranking.
Voice replication through AI Studio is not available in Illinois, Texas, EEA, UK, Switzerland, and India.Google, Gemini 3.8 text-to-speech says hello, footnote, September 23, 2026
05 — The decisionThe one decision #
A promotional price with a published end date is an invitation to prototype cheaply and a trap for anyone who quotes a client on it. The routing below assumes the January 1 rate is the real one.
Our earlier guide to Google's TTS models covers the Flash and Pro TTS line that came before this release. Teams that want the pricing and consent rules built into a content pipeline can look at how we run an AI content engine for clients.
06 — ConclusionThe promotional rate is real until December 31, and so is the doubling #
Prototype on today's price, put the January 1 price in every forecast, and clear the consent and geography rules before cloning anyone
Gemini 3.8 Flash TTS is a stable release with a documented price path. The work this week is arithmetic: measure minutes, multiply by the January rate, and decide whether Flash or Flash-Lite, Standard or Batch, fits the product. Voice replication is the feature to handle carefully, since its consent and geography rules are Google's, and yours to comply with.