Gemini 3.8 Live and Gemini 3.8 TTS: pricing, benchmarks and known issues

Gemini 3.8 Live is Google’s voice model for real-time spoken conversation and voice agents, released on 15 September 2026 alongside a reasoning variant, Gemini 3.8 Live Extended Thinking. Two text-to-speech models, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, became generally available on 22 September and were announced the next day. Live costs about $0.005 a minute for audio input and $0.018 a minute for audio output, and the TTS models start at $6 per million audio tokens until 31 December 2026. All prices are US list prices, checked with availability and leaderboard positions on 16 October 2026.

The answer in brief

The Gemini 3.8 voice models are worth testing now for voice agents, voiceovers and dubbing, as long as you check accents and voice consistency before anything reaches a customer. Extended Thinking is the strongest Google voice model for getting a task done and is cheaper per hour than the rivals that now outscore it. The TTS models are cheap until the end of the year, then double in price.

• Four models: Gemini 3.8 Live and Live Extended Thinking hold spoken conversations; Gemini 3.8 Flash TTS and Flash-Lite TTS read scripts aloud. All four are stable releases in the Gemini API.
• Quality: Extended Thinking launched at the top of the Artificial Analysis Speech to Speech Index with 82.6. On 16 October 2026 it was third, behind Grok Voice 3.0 and OpenAI’s GPT-Live-1 (Astra).
• Live price: about $0.005 a minute in and $0.018 a minute out. On Artificial Analysis’s fixed test that came to $0.84 an hour of input audio for Live and $3.50 for Extended Thinking, whose thinking is billed as text output.
• TTS price: $9 per million audio tokens on Flash TTS and $6 on Flash-Lite TTS until 31 December 2026, then $18 and $12. An hour of speech is 90,000 tokens.
• Against ElevenLabs: listeners rate ElevenLabs’ newest voices higher, but Gemini 3.8 Flash TTS costs under half the list price of Eleven v4 Turbo.
• UK: developers can use all four models, but voice replication in Google AI Studio is blocked in the UK.
• Businesses: Gemini 3.8 Live has been generally available in Gemini Enterprise since 24 September, with an optional talking video avatar, on US and EU endpoints. Extended Thinking is still in private preview there, and the TTS models are in preview on Google Cloud.
• Deadline: Google shuts down Gemini 3.1 Flash Live Preview, the Gemini 2.5 Flash native audio preview, Gemini 3.1 Flash TTS Preview and the Gemini 2.5 preview TTS models on 17 November 2026, according to its deprecations page.
• Known issues: developers report accents drifting back to American, voices changing mid-recording and flatter delivery than the 3.1 models. Google says it is looking into the consistency fault.

What are the Gemini 3.8 voice models?

Two of the models hold live conversations and two read scripts aloud. None carries a preview tag in the Gemini API, so Google treats all four as ready for production there.

The Live models take text, images, audio and video in, and answer in text and audio, with up to 131,072 input tokens and 65,536 output tokens according to the model page. Google’s deprecations page names Gemini 3.8 Live as the replacement for Gemini 3.1 Flash Live Preview and the Gemini 2.5 native audio models, and its release notes say Flash-Lite TTS is built to replace Gemini 3.1 Flash TTS Preview.

Gemini 3.8 Live is a different product from Gemini 3.8 Flash, Google’s text model of the same generation, with its own model ID and prices. Nor do these models compete with Gemini 4 Argon, Google’s newest top model, which writes text only.

Gemini 3.8 Live also has a visual mode for businesses. Gemini 3.8 Live with Live Avatar, launched on 24 September 2026, gives a voice agent a talking video face with lip-sync and facial expressions in all 97 languages. Companies choose from Google’s library of preset avatars, and a custom avatar made from a reference image needs approval through Google’s allowlist. Google launched it in Gemini Enterprise; neither announcement mentions the Gemini API or Google AI Studio. All the audio and video it generates carries an imperceptible SynthID watermark, according to Google Cloud.

What is the difference between Gemini 3.8 Live and Live Extended Thinking?

Gemini 3.8 Live answers straight away; Extended Thinking keeps talking while it works something out in the background. Plain Live reads camera or screen input in near real time, switches automatically between 97 languages mid-conversation and can run tool calls in the background while it chats, according to Google’s launch post. Its reasoning runs on a fixed latency profile, so you cannot turn it up.

Extended Thinking reasons and speaks at once. When a request needs planning or a slow system, it says something like “Checking flight options now”, fires off the tool calls and narrates its progress until the answer is ready. Google’s developer guide sets out the differences:

In practice, use plain Live when your systems answer in milliseconds and each question needs one reply. Use Extended Thinking when the caller would otherwise sit in silence while your back end works.

How good is Gemini 3.8 Live?

Gemini 3.8 Live Extended Thinking scores 82.6 on the Artificial Analysis Speech to Speech Index, the figure Google quoted at launch, and plain Live scores 76.0. On 15 September 82.6 was first place. A month later, rivals have overtaken it. Grok Voice 3.0 from SpaceXSI (formerly SpaceXAI), released on 14 October, now leads, and OpenAI’s GPT-Live-1, run with GPT-6 Astra at medium reasoning, is second. The index averages four tests: spoken reasoning, completing customer-service tasks with tools (τ-Voice), people’s preferences in blind conversations, and task success. Our Grok Voice 3.0 review compares the new leader with both rivals.

The two Google models split the work. Extended Thinking completes far more customer-service tasks than plain Live, 68.6% against 30.1%, and nearly twice the 37.7% of the model it replaces. Plain Live wins on feel: people in Artificial Analysis’s blind Speech Agent Arena prefer it to Extended Thinking, by an Elo rating of 1,088 against 993, and it starts speaking slightly sooner, 1.18 seconds against 1.35. It also has the lowest cost per hour of input audio of any model with an index score.

Two cautions apply. Artificial Analysis bases the GPT-Live-1 task scores on a single run where most models get three, so treat those as the noisier figures. Google’s other claims, 97.7% on Big Bench Audio and 35.1% on Sierra’s τ-Voice-banking test, come from its own launch post.

The table leaves out GPT-Live-1.1, which reached ChatGPT users from 15 October 2026. Developers cannot call it yet (OpenAI promises API access soon), and nobody has published scores for it.

How much does the Gemini 3.8 Live API cost?

The Gemini 3.8 Live API costs about $0.005 a minute for audio input and $0.018 a minute for audio output, and both Live models share one rate card. In tokens, that is $3.00 per million audio input tokens and $12.00 per million audio output tokens, according to Google’s pricing page. Text costs $0.75 in and $4.50 out per million tokens, and image or video input $1.00 per million. Unlike the TTS models, the Live models have no introductory rate and no scheduled rise.

The shared rate card hides a large gap in real bills. Extended Thinking’s background reasoning is charged as text output, so a conversation costs more the harder the model thinks. On Artificial Analysis’s fixed test, that made it about four times the hourly cost of plain Live. The thinking level (low, medium or high) is the obvious lever on that cost.

The free tier covers both models, but Google may use free-tier conversations to improve its products, so keep client data on the paid tier. Grounding answers in Google Search is free for 5,000 requests a month across Gemini 3 and newer models, then $14 per 1,000.

Which voices and languages does Gemini 3.8 Live support?

Gemini 3.8 Live understands and speaks 97 languages and notices when a speaker switches language mid-conversation. It speaks in Google’s prebuilt voices, such as Puck in the developer guide examples. Two limits matter for anyone choosing a voice. Custom voices made with the new Voice Design tool are documented for the two TTS models only, and a developer building a Spanish-language phone agent found they cannot yet be used in Live. Requests for a particular accent are also followed less reliably than on Gemini 3.1 Flash Live, as the British accent section below explains.

What’s new in Gemini 3.8 Flash TTS and Flash-Lite TTS?

Flash TTS is built for directed performance and Flash-Lite TTS for volume. Both take the same request format, so switching between them means changing one model ID. Google’s announcement and documentation list the main features:

• Voice design. Describe a voice in plain words, such as role, accent and character, and the model creates it. Saved voices last a year, with up to 200 per project, according to the voice design guide.
• A larger library. More than 2,000 ready-made voices in Google’s Extended Voice Library, including regional varieties such as Mexican Spanish, Quebec French and Scots English, on top of 30 prebuilt voices, according to Google Cloud.
• Voice replication. A copy of a real voice from a 10 to 30 second clip, gated by a spoken consent statement from the same adult speaker, according to the replication guide.
• Line-by-line direction. Control over pacing, emotion and accent, with tags for sounds such as laughs and sighs and for listener noises such as “mhm”.
• Scenes and long recordings. Two speakers from a single script, and long recordings that Google says keep the same voice for hours.

On benchmarks, Google says Flash TTS ranks first on Hume AI’s Voice Design Benchmark with 71.4 and first for accents with 60.8, and that Flash and Flash-Lite take the top two places on Hume’s Overall Quality Index. It also claims top places in blind Voice Arena tests in Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi; British English is not on that list. Every clip the Gemini audio models produce carries Google’s inaudible SynthID watermark, and replicated voices also carry C2PA content credentials.

How much does Gemini TTS cost?

Gemini 3.8 Flash TTS costs $9.00 per million audio output tokens and Flash-Lite TTS $6.00 until 31 December 2026, and both prices double on 1 January 2027. Text input is $0.50 per million for both, and the free tier covers both models. Audio is billed at 25 tokens a second, so an hour of speech is 90,000 tokens. The rates below are from Google’s pricing page; the hourly figures are AIWIZ’s arithmetic.

Even the 2027 rates undercut the $20.00 per million audio tokens that Gemini 3.1 Flash TTS Preview charges until it shuts down on 17 November 2026. Build any business case on the January prices, because those are the ones you will pay next year.

Is Gemini 3.8 Flash TTS better than ElevenLabs?

On sound alone, ElevenLabs is ahead; on price, Gemini wins easily. In the Artificial Analysis text-to-speech arena, where listeners choose between unlabelled clips, ElevenLabs’ Eleven v4 Turbo and Eleven v4 led on 16 October 2026. Gemini 3.8 Flash TTS ranked fifth and Flash-Lite TTS eighth.

Gemini 3.8 Flash TTS costs under half the list price of Eleven v4 Turbo and about a fifth of Eleven v4. The arena tests each provider’s stock voices, and Artificial Analysis gives Gemini’s rank a range of fourth to seventh, so treat small gaps as ties. For high-volume narration or dubbing, Gemini’s price is hard to ignore; for the voices listeners rate most natural, ElevenLabs still leads.

Can Gemini 3.8 TTS clone your voice in the UK?

Not in Google AI Studio. A footnote to Google’s announcement says voice replication there is unavailable in the UK, the EEA, Switzerland, India, Illinois and Texas. The footnote covers replication only, so voice design and the voice library still work.

Google’s replication guide for the API did not mention any regional limit on 16 October 2026, and the announcement does not say whether the AI Studio block also applies to API calls. Ask Google before you plan a UK product around a replicated voice. Where replication is offered, the API takes two recordings from the same adult speaker, the voice sample and Google’s consent statement read aloud, and checks that they match before it builds the voice.

Can Gemini speak with a British accent?

Yes, but how reliably depends on which product you use. In the Gemini app, Capella is the voice Google described as having a British accent before it dropped voice descriptions in May 2026, according to 9to5Google. In Gemini 3.8 Flash TTS, you can describe a British or regional accent with Voice Design, and the voice library includes regional varieties such as Scots English.

The Live API is the weak spot. A developer reported on 23 September that Gemini 3.8 Live ignores a system instruction to use a British accent, which Gemini 3.1 Flash Live had followed. Others in the thread describe the same drift back to American with South African and Scottish accents, and one team has moved speech output to a separate text-to-speech provider as a workaround. Google staff have asked for details privately; no public fix had been posted by 16 October 2026.

Is Gemini 3.8 Live available in the UK?

For developers, yes. The United Kingdom is on Google’s list of available regions for the Gemini API and Google AI Studio, so UK developers can use all four models today, and only voice replication in AI Studio is blocked. Business and consumer access is patchier, as Google’s launch posts and its Google Cloud release notes set out:

Google names no countries for the consumer features, so try Search Live and Gemini Live from a UK account first. For business deployments, Google lists US and EU endpoints for Live with Live Avatar and names no UK endpoint, so confirm where your audio and video will be processed.

Why does the Gemini 3.8 voice keep changing?

Because of a fault Google has acknowledged but not yet fixed. The reports below come from Google’s developer forum and have not been independently verified, although Google staff have replied in the threads on accents and voice consistency.

Voices change mid-recording. In a thread on Gemini 3.8 TTS, users describe sudden shifts in tone during two to ten minute recordings, including a female voice turning male and words added that were not in the script. One team measured a similar fault on Gemini 3.8 Live with the prebuilt Kore voice: about 0.1% of replies in audio-only sessions, rising to 2 to 4% when video was streamed in and a woman or child was speaking. On 15 October Google staff said they were still looking into it and would post an update once it was resolved.

Less personality. Builders of companion and assistant apps say the 3.8 voices sound flatter than 3.1 and give shorter, clipped replies. That fits a documented change: Google’s migration notes say affective dialogue has been removed from the API for Gemini 3.8 Live.

Accents slip back to American. Gemini 3.8 Live drifts away from requested accents, as described in the British accent section above.

What can UK marketing teams do with the Gemini 3.8 voice models?

The best first uses are jobs where a person can check the result before a customer hears it, or where a person picks up the hard cases. These are proposed uses to test, not results AIWIZ has measured.

Enquiry calls. Plain Live can answer opening-hours, pricing and “do you cover my area” questions and route the rest, at $0.84 an hour of caller audio on Artificial Analysis’s measure. Extended Thinking suits bookings that need your calendar or stock system. Tell callers they are speaking to an AI, hand over to a person on request, and check your call-recording and data protection arrangements before it goes live.

Voiceovers for video and social. Flash TTS takes stage directions line by line and can stage a two-person scene from one script, which suits explainer videos and podcast-style trailers. Google is rolling Flash-Lite TTS out in Google Vids, its video editor.

Localisation. With 130 languages on Flash TTS and 101 on Flash-Lite TTS, dubbing a video or reading a landing page aloud in a client’s export markets costs 0.9 to 1.4 US cents a minute of finished audio until December, and twice that from January.

A consistent brand voice. Voice Design can create one voice from a written description and keep it for a year. Until the consistency fault is fixed, listen to every long recording in full before it is published.

Video avatars. Live Avatar puts a talking face on a website or kiosk assistant, for jobs such as Google’s own example of checking in a hotel guest, or a guided product walkthrough. It needs Gemini Enterprise, and a custom avatar of a real spokesperson needs Google’s approval as well as that person’s written consent.

Voice search. Search Live runs on Gemini 3.8 Live, so people can now ask Google a question out loud and keep following up. Being cited in spoken answers takes the same groundwork as typed AI answers: clear, sourced replies to the questions customers actually ask. Our answer engine optimisation service is built around that groundwork.

A fair pilot takes five steps:

1. Pick one script or call flow. A product video voiceover or the ten questions your reception team answers most often both work, because the right answer is known.
2. Write the checklist first. Accent, pronunciation of your brand and product names, facts it must not invent, and when it must hand over.
3. Run it twenty times. Voice drift and accent slips are intermittent, so one good take proves little.
4. Listen in full. Note every change of voice, tone or accent, and every added word.
5. Compare cost per approved minute. Count model charges, review time and retakes, at January prices as well as today’s.

What changes for developers moving to the Gemini 3.8 voice models?

Moving to plain Gemini 3.8 Live is close to a drop-in swap; Extended Thinking and the TTS models need code changes. The quickest way to try the Live models is the Live playground in Google AI Studio, and Google names LiveKit, Pipecat, Agora and Vercel among the platforms that already support them.

• Gemini 3.8 Live: change the model string to gemini-3.8-live and remove any thinking level setting. Function calls now run in the background by default, and affective dialogue has gone.
• Extended Thinking: declare every function as non-blocking, set the thinking level, and drive your interface from the interaction status rather than the end of each turn, because the model can speak several times within one task.
• TTS from Gemini 3.1 Flash TTS Preview: the input text is now read strictly as a verbatim transcript, so delivery notes belong in separate speech metadata, per the Flash-Lite TTS model page. Unary requests return WAV with a header by default.
• Long prompts: replace long “Audio Profile” prompts with a saved designed voice; Philipp Schmid warns that long profiles now cause voice drift. Each TTS request takes at most 8,192 input tokens, according to Google Cloud, so split long scripts.

Plan the move now. Google’s deprecations page shuts down Gemini 3.1 Flash Live Preview and the Gemini 2.5 Flash native audio preview on 17 November 2026, with Gemini 3.8 Live as the replacement. Gemini 3.1 Flash TTS Preview and the Gemini 2.5 Flash and Pro preview TTS models close the same day, with the two Gemini 3.8 TTS models as replacements. Teams that rely on the 3.1 model’s steadier accents have about a month to test 3.8 Live or move speech output to a separate text-to-speech model.

Should you switch to the Gemini 3.8 voice models?

For most voice work, test them now and switch job by job. The case is strongest on price, and the risk sits in accents and consistency.

If you are building a voice agent that books, diagnoses or looks things up across several systems, test Extended Thinking: it is Google’s strongest voice model for completing tasks and costs less per hour than the two models that now outscore it. For quick conversation with simple tools, plain Gemini 3.8 Live has the lowest hourly cost of any model on the Artificial Analysis index, and people prefer it in blind comparisons. For narration, dubbing and read-aloud audio, Flash-Lite TTS at $6 per million audio tokens is hard to beat until 31 December. If your product depends on a fixed British accent or one voice across long recordings, run the listening pilot above now, because the Gemini 3.1 Flash Live preview that held accents better shuts down on 17 November 2026.

To try a voice agent or a voiceover workflow on your own material before you commit, talk to AIWIZ. We can run the five-step pilot above on one script or call flow, and our AI adoption services cover rolling out whichever set-up comes out ahead.