Gemini 3.8 Flash TTS is Google’s new flagship text-to-speech model, released on 23 September 2026 alongside a cheaper sibling, Gemini 3.8 Flash-Lite TTS. Both went live the same day in the Gemini API and Google AI Studio as generally available models rather than previews, and Google calls them its most expressive audio generation models yet. The headline features are voices designed from a text prompt, voices replicated from about 30 seconds of audio, and line-by-line direction of how a script is performed.
The launch post, written by group product manager Leland Rechis and Alan Cowen, director of research science on the Gemini audio team, leads with benchmark wins. The developer documentation and price list tell a more useful story: introductory prices that double on 1 January 2027, a two-speaker limit, and a voice replication feature that is switched off for users in the UK. Our earlier explainer on Gemini 3.1 Flash TTS covers the model this one replaces.
This article covers what each model is for, what developers get in AI Studio and the API, what the pricing is and when it changes, who ran the benchmarks, and which restrictions matter most for UK and EU teams.
Table of contents
- What Google Released: Gemini 3.8 Flash TTS and Flash-Lite TTS
- What Gemini 3.8 Flash TTS Can Do
- Voices and Languages in Gemini 3.8 Flash TTS
- Gemini 3.8 Flash TTS Pricing and the 1 January Price Rise
- How to Estimate What Gemini 3.8 Flash TTS Will Cost
- The Benchmark Claims Behind Gemini 3.8 Flash TTS
- Safety, Consent and Watermarking
- What Gemini 3.8 Flash TTS Means for UK and EU Developers
- Where Gemini 3.8 Flash TTS Fits in Google’s Voice Stack
- Gemini 3.8 Flash TTS Compared With Earlier Gemini Speech Models
- Building a Voice Agent With Gemini 3.8 Flash TTS
- Limitations to Plan Around With Gemini 3.8 Flash TTS
- How to Choose Between Flash TTS and Flash-Lite TTS
- Frequently Asked Questions About Gemini 3.8 Flash TTS
- References
What Google Released: Gemini 3.8 Flash TTS and Flash-Lite TTS
Google shipped two speech models with deliberately different jobs, and choosing the wrong one costs either quality or money.
Two models, two jobs
Gemini 3.8 Flash TTS is positioned for “deep creative direction and character design” across gaming, audiobooks, podcasts and interactive media. Gemini 3.8 Flash-Lite TTS is built for “high-volume, cost-efficient scale”, optimised for dubbing, audio content generation and expressive voice agents.
Generally available on day one
The Gemini API documentation lists gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts as generally available and production-ready. The older gemini-3.1-flash-tts-preview and gemini-2.5-pro-preview-tts remain available as previews but are superseded.
Where each model appears
Gemini 3.8 Flash TTS is available today to developers in the Gemini API and AI Studio, and to general users in Gemini Notebook. Flash-Lite is in the API and AI Studio, and powers voice in Google Vids. Both are listed as “coming soon” to the Gemini Enterprise API.
Inputs and outputs
Both models take text only and return audio only. Unite.AI, citing the model card, reports text input of up to 8,000 tokens and audio output of up to 64,000 tokens per request.
| Feature | Gemini 3.8 Flash TTS | Gemini 3.8 Flash-Lite TTS |
|---|---|---|
| Model ID | gemini-3.8-flash-tts | gemini-3.8-flash-lite-tts |
| Built for | Creative direction, characters | High volume, low cost |
| Languages in the API docs | 130 | 101 |
| Audio output price to 31 Dec 2026 | $9 per 1M tokens | $6 per 1M tokens |
| Consumer product | Gemini Notebook | Google Vids |
| Status | Generally available | Generally available |
What Gemini 3.8 Flash TTS Can Do
Most of the new capability is about control: over what a voice sounds like and over how each line is delivered.
Voice design from a prompt
Google says developers can scale “from 30 original voices to infinite library via generative voice design”, describing a new voice in plain language. Voice design is as much a natural language processing problem as an audio one, because the model has to turn a written description of age, accent and character into a consistent sound.
Voice replication from 30 seconds
Gemini 3.8 Flash TTS can replicate a specific voice from roughly 30 seconds of audio. Replication requires a recorded verbal consent statement from the voice owner, which Google checks against the reference speaker before the voice is created.
Line-by-line direction
Scripts can carry natural cues for acting, pacing and dialect shifts on individual lines. Google also claims long-form generation that holds quality across hours of audio, which matters for audiobooks more than any single benchmark does.
Vocal bursts and backchanneling
Inline tags add non-verbal sounds, such as <laugh>, <sigh>, <gasp>, <breath>, <cough> and <short pause>. The documentation’s example is Wait... <short pause> did you hear that? <sigh>. Backchanneling adds the small listening noises that make a two-person conversation sound real.
Two-speaker scenes
Gemini 3.8 Flash TTS stages two speakers natively in one request. The documentation caps that at two speakers using prebuilt voices; custom voices must be generated separately and joined afterwards.
Voices and Languages in Gemini 3.8 Flash TTS
Voice supply has grown more than sixtyfold, and language coverage now differs between the two models.
Thirty prebuilt voices
The API ships 30 studio voices with names such as Zephyr, Puck, Charon, Kore, Fenrir and Aoede. These are the voices the two-speaker mode can use directly.
More than 2,000 production voices
An Extended Voice Library adds hundreds more voices, filterable by language, accent, gender, pitch and persona, and Google puts the total at more than 2,000 production-ready voices.
Languages
The launch post says more than 100 languages and dialects, including Mexican Spanish, Quebec French and Scots English. The API documentation is more precise: 130 languages for Gemini 3.8 Flash TTS and 101 for Flash-Lite. If you need a less common language, check the list for the specific model.
Saving custom voices
Designed and replicated voices are returned with persistent voice_ identifiers, capped at 200 per project and kept for one year. Stateless voicekey_ identifiers last seven days. Google says voice remixing, adjusting timbre, pitch, pace and accent, is coming soon.
The prebuilt set is a tiny fraction of what is now available, and the per-project cap on saved custom voices sits between them.
Gemini 3.8 Flash TTS Pricing and the 1 January Price Rise
The price list carries the most important detail in the launch, and the launch post does not mention it.
Introductory prices
Gemini 3.8 Flash TTS costs $0.50 per million text input tokens and $9.00 per million audio output tokens “through December 31, 2026”. From 1 January 2027 those become $1.00 and $18.00.
Flash-Lite prices
Gemini 3.8 Flash-Lite TTS costs $0.50 input and $6.00 output until the end of 2026, then $1.00 and $12.00. Both models have a free tier.
Batch halves the cost
Batch processing halves every rate. Flash TTS batch output is $4.50 now and $9.00 from January; Flash-Lite batch output is $3.00 now and $6.00 from January.
Against the model it replaces
The gemini-3.1-flash-tts-preview model costs $1.00 input and $20.00 output. Today Gemini 3.8 Flash TTS output is 55 per cent cheaper than that. From January the gap shrinks to 10 per cent. Flash-Lite goes from 70 per cent cheaper to 40 per cent cheaper.
| Model, per 1M tokens | Input to 31 Dec 2026 | Output to 31 Dec 2026 | Output from 1 Jan 2027 | Batch output now |
|---|---|---|---|---|
| Gemini 3.8 Flash TTS | $0.50 | $9.00 | $18.00 | $4.50 |
| Gemini 3.8 Flash-Lite TTS | $0.50 | $6.00 | $12.00 | $3.00 |
| Gemini 3.1 Flash TTS preview | $1.00 | $20.00 | No change listed | $10.00 |
The January rise takes the new flagship almost back to the old preview’s price.
How to Estimate What Gemini 3.8 Flash TTS Will Cost
The price list is per token, and that creates a budgeting problem the launch materials do not solve.
Priced per token, used in minutes
Speech is consumed in minutes and hours, but Gemini 3.8 Flash TTS is billed per million audio output tokens. To budget an audiobook or a month of voice-agent replies, you need to know how many tokens a minute of speech produces.
The missing conversion rate
Google’s token documentation says audio input uses 32 tokens per second. It gives no equivalent rate for generated speech, so there is no published way to turn the $9 output price into a cost per minute of Gemini 3.8 Flash TTS audio.
Measure it yourself
The practical fix is to generate a representative sample, such as ten minutes of your own script at your chosen pace, and read the token counts the API returns. Multiply up from there, and repeat the test on Flash-Lite, because the two models may not produce the same number of tokens for the same speech.
Build the January price into the model
Whatever rate you measure, cost Gemini 3.8 Flash TTS at $18 per million output tokens and Flash-Lite at $12 from 1 January 2027. A business case that only works at the introductory price is a business case with a three-month life.
| What to measure | Why it matters | How to get it |
|---|---|---|
| Output tokens per minute of speech | Converts the token price into a cost per minute | Token counts on a ten-minute test |
| Input tokens per script | Direction cues and tags add input tokens | Token counts on your real scripts |
| Share of work suitable for batch | Batch halves every rate | Split live replies from pre-generated audio |
| Monthly volume after 1 January 2027 | Output prices double | Forecast at the 2027 rate |
The Benchmark Claims Behind Gemini 3.8 Flash TTS
Google’s benchmark results are strong. The context around one of them deserves a sentence the launch post does not include.
Hume AI Voice Design Benchmark
Google reports that Gemini 3.8 Flash TTS took the top overall position on Hume AI’s Voice Design Benchmark with a score of 71.4, and led accent modelling at 60.8.
Hume AI Overall Quality Index
On Hume AI’s Overall Quality Index, Google says Flash TTS ranks first and Flash-Lite TTS second.
Voice Arena
In Voice Arena’s blind human-preference leaderboards, Google reports top positions in Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi.
Who ran the tests
Alan Cowen, a named author of Google’s launch post, founded Hume AI. Google DeepMind hired him in a licensing deal with Hume AI announced in January 2026. That does not make the Hume scores wrong, but it means the two Hume benchmarks are not independent in the usual sense. The Voice Arena results, based on blind human preference, are the stronger evidence for Gemini 3.8 Flash TTS.
What the model card says
Unite.AI reports that the model card’s frontier safety assessment found no meaningful new capabilities compared with Gemini 3.7 Flash, and lists hallucinations and occasional slowness or timeouts as known limitations.
Safety, Consent and Watermarking
Voice cloning is the part of text-to-speech with the clearest route to harm, and Google has put several controls around it.
SynthID on every output
All audio generated by both models carries SynthID, Google’s imperceptible watermark for identifying AI-generated content.
C2PA credentials for replicated voices
Replicated voices also carry C2PA content credentials, the industry standard for recording where a piece of media came from.
Consent before replication
The voice owner must record a verbal consent statement, and Google matches that recording to the reference speaker before creating the voice.
Regions where replication is unavailable
Voice replication is not available in Illinois, Texas, the European Economic Area, the United Kingdom, Switzerland or India. Google does not give a reason, but several of these jurisdictions have specific biometric or data protection rules.
| Control | Applies to | What it does |
|---|---|---|
| SynthID watermark | All generated audio | Marks audio as AI-generated |
| C2PA credentials | Replicated voices | Records provenance of the media |
| Verbal consent match | Voice replication | Checks the owner agreed |
| Regional block | Replication in IL, TX, EEA, UK, CH, India | Feature not offered |
What Gemini 3.8 Flash TTS Means for UK and EU Developers
For teams in the UK and Europe the launch is useful, but not in quite the way the headline suggests.
Voice design works; replication does not
UK and EEA developers can design voices from prompts and use the full prebuilt and extended libraries. They cannot replicate a real person’s voice, which rules out the most-publicised feature for dubbing an existing presenter.
Telephony-ready output
Output defaults to 24 kHz WAV, but the API also offers 16 kHz and 8 kHz sample rates and mu-law and A-law encodings, the formats phone systems use. That makes Gemini 3.8 Flash TTS a realistic option for contact centre voice agents without a conversion step.
Budget for January
Any cost model built on today’s prices is wrong from 1 January 2027. Model the doubled rate from the start, and treat the next three months as a discount.
Governance first
Synthetic voices of staff or customers raise consent and disclosure questions. Our AI strategy team helps organisations set policy before deployment rather than after.
Where Gemini 3.8 Flash TTS Fits in Google's Voice Stack
The speech models sit alongside Google’s conversational and consumer voice products, not in place of them.
Consumer surfaces
Gemini 3.8 Flash TTS voices Gemini Notebook, and Flash-Lite voices Google Vids. Our coverage of Gemini Notebook’s voice mode describes the first of those.
Partner platforms
Google names Agora, LiveKit, Pipecat and Vercel as developer platforms, and Figma, HeyGen, Linguana, Wondercraft, 99.co, Ollang, Spoken and Transforms.AI among early partners.
The wider Gemini 3.8 family
The text model Gemini 3.8 Flash, which we covered in our Agent Studio report, is a separate model. The TTS models share the version number but do a single job: turning text into speech.
Gemini 3.8 Flash TTS Compared With Earlier Gemini Speech Models
The new models replace two previews that are still listed, which matters for anyone who built on them.
Gemini 2.5 Pro Preview TTS
The oldest model still in the documentation remains a preview and is marked as superseded. There is little reason to start anything new on it.
Gemini 3.1 Flash TTS preview
The model covered in our April explainer costs $1.00 per million input tokens and $20.00 per million output tokens. Gemini 3.8 Flash TTS undercuts it by 55 per cent on output until the end of 2026.
What general availability changes
A generally available model is the one Google positions as production-ready, while previews carry fewer guarantees about stability. Moving from the 3.1 preview to Gemini 3.8 Flash TTS is therefore a stability upgrade as well as a quality one.
Migrating from the preview
Changing the model identifier is the easy part. Re-listen to every voice you depend on, because a new model can change how an existing prebuilt voice sounds, and re-check your cost model against the new token prices.
Building a Voice Agent With Gemini 3.8 Flash TTS
Voice agents are where Flash-Lite is aimed, but most of the engineering sits around the speech model rather than in it.
The usual pipeline
A voice agent listens with speech recognition, decides what to say with a language model, and speaks with a text-to-speech model. Gemini 3.8 Flash TTS and Flash-Lite fill only the last step, so responsiveness depends on the whole chain.
Streaming output
The API streams headerless raw audio at 24 kHz by default, which lets playback start before a full reply has been generated. On a live call that matters more than any quality benchmark.
Partners do the plumbing
Google names Agora, LiveKit, Pipecat and Vercel as developer platforms. Frameworks of that kind handle turn-taking, interruptions and telephony, which is why they appear in the launch rather than as an afterthought.
Keep turns short and tagged
Agents speak in short turns. A backchannel sound or a short pause before a difficult answer makes a synthetic voice sound less abrupt, and the tags add almost nothing to the output bill.
Limitations to Plan Around With Gemini 3.8 Flash TTS
None of these stops a project, but each has caught teams out with earlier speech models.
Text in, audio out only
The TTS models accept text and return audio. Any agent that needs to understand speech still needs a separate model for listening.
Two speakers per request
Scenes with three or more voices, such as a drama or a panel podcast, must be generated in parts and stitched together, which adds editing work and cost.
Voices expire
Saved custom voices last one year and stateless voice keys seven days. Keep the design prompts you used so a voice can be recreated if it lapses.
Known model limitations
Unite.AI reports that the model card lists hallucinations and occasional slowness or timeouts. In speech, a hallucination can mean words that were never in the script, so check long outputs before publishing them.
How to Choose Between Flash TTS and Flash-Lite TTS
The choice comes down to how much direction a script needs and how many hours of audio you will generate.
Choose Gemini 3.8 Flash TTS for performance
Audiobooks, games, podcasts and any content where a listener notices delivery belong on Gemini 3.8 Flash TTS. It also covers 130 languages against Flash-Lite’s 101.
Choose Flash-Lite for volume
Dubbing, notifications and voice agents that speak thousands of short replies suit Flash-Lite, at two thirds of the output price.
Use batch where latency does not matter
Overnight generation of audio content at batch prices costs half as much on either model.
Prototype in AI Studio first
AI Studio lets you compare both models on the same script before writing any code, which is the cheapest way to hear the difference.
Frequently Asked Questions About Gemini 3.8 Flash TTS
What is Gemini 3.8 Flash TTS?
Google’s text-to-speech model for expressive, directed speech, released on 23 September 2026 in the Gemini API and Google AI Studio.
How much does Gemini 3.8 Flash TTS cost?
$0.50 per million input tokens and $9.00 per million audio output tokens until 31 December 2026, then $1.00 and $18.00.
Can I clone a voice in the UK?
No. Voice replication is not available in the United Kingdom, the EEA, Switzerland, India, Illinois or Texas.
How many speakers can one request include?
Two, using prebuilt voices. Custom voices must be generated separately and combined.
Is the audio watermarked?
Yes. All output carries SynthID, and replicated voices also carry C2PA credentials.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
