Google's new voice models undercut everyone — but the headline feature is unavailable in the EU, and the price doubles on 1 January

What shipped

Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS reached general availability on 23 September 2026, logged in the Gemini API release notes the same week and replacing Gemini 3.1 Flash TTS Preview. Both are available today in the Gemini API and Google AI Studio, with Gemini Enterprise listed as coming soon. Flash TTS also reaches everyone through Gemini Notebook; Flash-Lite TTS through Google Vids.

The capability list is genuinely broad. Both models cover 100-plus languages and dialects — Google names Mexican Spanish, Quebec French, Scots English, Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic and Hindi among them — and ship with over 2,000 production-ready voices. Beyond the library there are three paths to a voice: design one from a natural-language description, replicate one from a 30-second sample, or remix timbre, pitch, pace and accent, which Google lists as coming soon rather than shipping.

The direction controls are the part that will matter most to anyone producing long-form audio. Scripts take line-by-line stage directions in natural language, generation runs to hours of continuous audio, two-speaker scenes can be staged in a single pass, and the models handle vocal bursts and backchannelling through inline markup such as <laughs> and |mhm|.

The rate card, and the date attached to it

Per million tokens, from Google’s own pricing page. The per-hour column follows from the audio meter: 25 tokens per second, so 90,000 tokens per hour.

Against the model it replaces, this looks like a large generational cut. Gemini 3.1 Flash TTS Preview charged $1.00 and $20.00, which is $1.80 for that same hour. Flash TTS at 81 cents is 55% cheaper.

Except the pricing page carries a second line for every one of those rates: through December 31, 2026. On 1 January 2027, Flash TTS goes to $1.00 and $18.00, Flash-Lite to $1.00 and $12.00. An hour of Flash TTS audio becomes $1.62 — against the $1.80 its predecessor charged. The durable generational saving on the audio meter is 10%. The other 45 points are a promotion with a date on it.

That date should look familiar. It is the same 31 December 2026 expiry that the Gemini Flash text models carry, where Gemini 3.8 Flash launched at the identical introductory rate, with the identical expiry, that 3.7 Flash had three weeks earlier. The pattern holds on the speech line: the discount follows the calendar, not the model. Standardising on the newest TTS model does not buy a single extra day of runway, and neither does waiting for the next one.

Where the price stops winning

Normalised to characters — the unit the rest of this market sells in — the current rates are about $33 per million characters for Flash TTS and $22.10 for Flash-Lite. Cartesia Sonic 3.6 is $49. ElevenLabs Eleven v3 is $100, at $0.10 per thousand characters.

On those numbers Google is the obvious buy, and the quality gap points the same way: on the Artificial Analysis Provider Voice Arena, Gemini 3.8 Flash TTS leads Eleven v3 by 93 Elo points, a margin with no overlap in the confidence intervals.

Run the same table at January rates and it reorders. Flash TTS doubles to roughly $66 per million characters — above Cartesia’s $49. And Cartesia is not a weaker model: on that same Voice Arena, Sonic 3.6 holds rank 1 at 1,273 Elo against Flash TTS at 1,260, with ±17-point intervals on both. Google’s model is not beating it now; in January it would also cost a third more.

Flash-Lite is the version that survives the reset. At about $44.20 per million characters it stays under Cartesia, and at rank 6 with 1,235 Elo it is a real step down in quality but not a collapse. For high-volume, less voice-critical work — the content pipelines where an hour of narration is a commodity input — that is the line item to model.

The benchmark provenance problem

Google’s announcement leads with rankings: #1 on Hume AI’s Voice Design Benchmark at 71.4, and the top two places on Hume’s Overall Quality Index.

Those are Google-reported results about Google models, which is ordinary for a launch. The part that is not ordinary is who maintains the benchmark. Alan Cowen, Hume AI’s founder, is credited on this release as a Director of Research Science at Google DeepMind, having joined through a January licensing deal that moved Hume engineers to Google while leaving Hume free to serve other labs.

Nothing here suggests the benchmark was manipulated. But a vendor announcing a first-place finish on a yardstick built by a company whose founder now works for that vendor is not an independent result, and it should not be read as one. The independent reading is the blind-preference Voice Arena, and it says something more modest: statistical tie for first, not a clear win. This is the same discipline that applies to independently repriced intelligence indices — the leaderboard a vendor chooses to headline is itself a claim.

Consent, watermarking and the EU exclusion

Every generated clip carries a SynthID watermark, and voice replication output additionally carries C2PA credentials. That matters for the same reason it mattered when Google shipped Lyria 3.5 with SynthID on every output: Article 50(2) of the EU AI Act puts a machine-readable marking duty on the provider, and Google meeting it before the file reaches you removes one compliance question from your build. It does not remove the disclosure duties that sit on you as a deployer — the same split that applies to watermarked text output.

To replicate a voice, Google requires a verbal consent recording from the voice owner that matches the reference speaker. That is a meaningful gate, and it is more than most of this market asks for.

It is also, for a large share of readers, beside the point. Voice replication is unavailable in the EEA, the United Kingdom, Switzerland, India, Illinois and Texas. Google built the consent verification, added C2PA on top, and still did not ship the feature into those jurisdictions. That decision tells you how the compliance exposure is being priced: not as something a consent recording clears, but as something worth forgoing the market over.

For an EU buyer the practical consequence is simple. The feature the launch coverage is organised around is not on your menu. The voice library, voice design, stage directions and long-form generation all are — and those, not cloning, are what most production pipelines actually run on.

What to do with this

If you are generating narration at volume, Google is now the cheapest credible option in the AI audio tool field, and Flash-Lite is the version whose economics survive January. Cost the workload at the post-reset rates before you commit; the promotional numbers are not what you will be paying by the time a pipeline built this quarter is running at scale.

If you are building conversational voice rather than generating audio files, this is the wrong meter to compare. OpenAI’s GPT-Live-1 bills wall-clock seconds rather than audio tokens, which changes the cost of silence, hold queues and thinking pauses entirely. And if the workload is speech going the other direction, the transcription price floor is a separate calculation again.

If your entity or your users sit in the EEA or the UK, treat voice replication as unavailable and design without it. The rest of Gemini’s speech stack is fully available — and on the evidence of the independent leaderboard, the prebuilt and designed voices are where the quality claim actually holds up.

Questions or corrections? Email Pick Right. Want the full list? See all news.