Gemini 3.8 Live vs Gemini 3.8 TTS: A Conversation Loop and a Reading Voice Share a Generation Name

Gemini 3.8 Live is a speech-to-speech model that shipped on September 15, 2026 as gemini-3.8-live and gemini-3.8-live-extended-thinking. "Gemini 3.8 TTS" is not a model at all — it is the umbrella term that launch coverage, including Goog​le's own post title, uses for the pair of text-to-speech endpoints that arrived on September 23, 2026 as gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. So this is not the usual versus page, where two products fight over one job. It is a page about two jobs that share a generation number, and about the specific mistake people make when they treat the words "Live" and "TTS" as two flavours of the same thing.

The fork the shared generation number hides

Goog​le's speech-generation documentation draws the line itself, and the sentence is easy to skim past: "The TTS capability differs from speech generation provided through the Live API." The same passage explains why, and the reason is structural rather than a matter of quality. The Live API is built for "interactive, unstructured audio, and multimodal inputs and outputs." TTS through the Gemini API "is tailored for scenarios that require exact text recitation with fine-grained control." A later note makes the input shape explicit: the TTS models take text only and return audio only.

That single constraint — text in, audio out, no audio input — is the whole comparison. A Gemini 3.8 Live model hears the person talking to it. A Gemini 3.8 Flash TTS or Flash-Lite TTS model never hears anyone; it reads a string and performs it. Everything else follows from there: the benchmarks that exist for one are meaningless for the other, the pricing is metered on different units, and the failure modes are not comparable at all.

This matters because the two use cases look identical from the outside. A voice agent and a narration pipeline both end up emitting human-sounding speech. Only one of them can be interrupted.

Where "Gemini 3.8 TTS" actually resolves

Open the supported-models table for the Gemini API's speech generation and you will find four TTS entries and no entry called Gemini 3.8 TTS. Two of them are this generation: Gemini 3.8 Flash TTS, model ID gemini-3.8-flash-tts, with 130 languages, and Gemini 3.8 Flash-Lite TTS, model ID gemini-3.8-flash-lite-tts, with 101 languages. The other two — Gemini 3.1 Flash TTS Preview and Gemini 2.5 Pro Preview TTS — are the preview generation that Flash-Lite replaces.

So when someone says "Gemini 3.8 TTS" they are usually pointing at the Flash tier, because that is the one the launch post leads with and the one the Arena board ranks highest. When they say it about a live voice agent, they are pointing at nothing, because the thing they want is on a different endpoint with a different name and a different input contract.

There is a third collision worth naming, because it produces the worst advice. Gemini 3.8 Live is not a text-to-speech model with extra features. It is a speech-to-speech model, and the reason it costs more per unit is that it is doing more work per unit: listening, reasoning mid-utterance, handling interruptions, and in the Extended Thinking variant, deliberating before it answers.

The only number that genuinely compares

Quality scores do not transfer between these two families; there is no single board that scores a narrator and a conversationalist on the same axis. Money does transfer, once you pick the unit honestly, and the honest unit for anything voice is the audio-hour.

• Metered in — Gemini 3.8 Live bills per minute of audio, $0.005 per minute in and $0.018 per minute out vs Gemini 3.8 Flash TTS bills per million audio tokens, $9.00 through 31 December 2026 and $18.00 after
• Metered out — Gemini 3.8 Live two-way audio runs about $1.38 for an hour of simultaneous input and output at list, before any prompt caching vs Gemini 3.8 Flash TTS is a one-way stream, and one million characters of finished narration comes to $16.5 by Artificial Analysis's normalisation
• Text input — Gemini 3.8 Live bills text and audio on separate lines, $0.75 per million text tokens in and $4.50 out vs the TTS endpoints bill text input at $0.50 per million tokens
• Flash-Lite tier — Gemini 3.8 Live has no cheaper sibling; the choice is Live or Extended Thinking vs Gemini 3.8 Flash-Lite TTS lists at $6.00 then $12.00 per million audio tokens, with Artificial Analysis at $11.0 per million characters
• Input shape — Gemini 3.8 Live audio, video and text in, audio out vs the TTS endpoints text in, audio out

The Live figure is our arithmetic from Goog​le's published per-minute rates, not a Goog​le-published hourly number. The character figures are Artificial Analysis's normalisation and are the ones to quote when you are comparing synthesis tiers. Note that Artificial Analysis's own Speech to Speech board carries a separate cost-per-hour-of-input-audio column for the Live models — around $3.50 for Gemini 3.8 Live and $0.84 for the Extended Thinking variant — which is a different normalisation again and should not be lined up against the per-minute rate card as though the two were the same measurement.

What breaks when you pick the wrong one

The failure modes are instructive because they are so unalike.

Call a TTS endpoint when you needed a conversation loop and nothing errors. The request returns valid audio. You get a fluent, well-read response to a fixed string, and then silence — because there was never anything listening. Teams discover this when they build their first barge-in test and find that the "user" has to wait for the full clip to finish before the next turn can begin. Retrofitting a conversation onto a synthesis endpoint means adding speech recognition, a turn-taking policy and an interruption mechanism around it, at which point you have rebuilt a cascade and you are paying for two models instead of one.

Call a Live endpoint when you needed narration and you pay for capability you do not use. A Live model is metered in both directions, because it expects both directions. For an audiobook or a training-video voiceover, the input side is a script that never changes and the model never needs to hear anything. You are buying a conversation loop to read a document aloud.

The one case where the choice is genuinely hard is the cascade: recognition, then reasoning, then synthesis. That architecture is not obsolete, and for a lot of production systems it is still the right one, because it lets you swap each stage independently and pick a different vendor for each. It costs you latency and it costs you the prosody that a single end-to-end model keeps across a turn boundary. The trade is real in both directions.

Where the route already exists

OrcaRouter does not carry the Gemini 3.8 Live endpoints or the 3.8 TTS endpoints, and that is worth saying before anything else about routing, because the opposite is easy to assume from a catalogue this wide. What the catalogue does hold is the rest of the family — google/gemini-3.8-flash among 28 Goog​le models and 200-plus models overall, from one key at provider list price with no markup.

That matters specifically for the cascade. If your architecture is recognition, then reasoning, then synthesis, the reasoning stage is the one where a single key across many vendors pays off, because it is the stage you swap most often and the stage where a vendor price cut should reach your bill the same day rather than at the next contract renewal. It is also the stage where automatic failover earns its keep: a cascade has three places to break, and the middle one is the cheapest to make redundant.

A short decision rule

If the model has to respond to something it did not already have as text, you want Gemini 3.8 Live, and you should budget on Goog​le's per-minute audio rates and expect to think about which of the two variants you need. If the text is already written and the job is to perform it, you want Gemini 3.8 Flash TTS or Flash-Lite TTS, and you should budget per million characters and choose the tier on language coverage and on whether the quality gap is worth the price gap.

And if you are being asked to compare "Gemini 3.8 Live and Gemini 3.8 TTS" as though they were two products, the first useful thing you can do is ask which endpoint. The name on the brief does not resolve to one, and the answer changes the architecture rather than just the invoice.