Gemini 3.8 Flash TTS vs Gemini 3.8 TTS: One of These Names Is Not a Model

There is no model called Gemini 3.8 TTS. The vendor's speech-generation documentation lists exactly two text-to-speech endpoints in this generation — gemini-3.8-flash-tts, which the docs name Gemini 3.8 Flash TTS, and gemini-3.8-flash-lite-tts, named Gemini 3.8 Flash-Lite TTS — and both shipped on September 23, 2026. "Gemini 3.8 TTS" is the umbrella term the launch coverage uses for the pair, including in the vendor's own post title. So this is not a matchup between two products, and anyone who tells you it is has not opened the docs.

What the query is actually pointing at is a real decision, and a more interesting one than the naming suggests: which member of the family should you call? On the independent board the two are inside each other's error bars, and one of them costs a third less. That is the comparison worth making, so let us make it properly — and then deal with the third name that keeps getting folded into the same conversation.

What each name resolves to

The distinguishing test is not what a search result says. It is whether the string resolves to an endpoint with a rate card and a leaderboard row, and on that test the two names separate cleanly.

• Gemini 3.8 TTS — no model identifier, no supported-models entry, no published price, no row on Artificial Analysis. It names a launch and a product line, not something you can call.

• Gemini 3.8 Flash TTS — model ID gemini-3.8-flash-tts, 130 languages, generally available in the Gemini API, second on the Artificial Analysis Provider Voice Arena.

• Gemini 3.8 Flash-Lite TTS — model ID gemini-3.8-flash-lite-tts, 101 languages, generally available, sixth on the same board, and described by Google as the workhorse replacement for the preview model it supersedes.

Both real endpoints take the same API schema and prompting format, which is why the choice between them is a configuration value rather than an architectural one. The mistake to avoid is treating the umbrella as a third option and waiting for it to appear; there is nothing to wait for.

The decision the umbrella name is hiding

Once the naming is settled, the question a purchaser actually has is whether the more expensive tier is worth it. The quality case for Flash TTS over Flash-Lite TTS is thinner than the price difference implies, and the board says so more clearly than any review.

On the Artificial Analysis Provider Voice Arena, captured for this article on September 24, 2026, Gemini 3.8 Flash TTS is ranked second with an Elo of 1,260 and a 17-point interval, measured over 1,999 samples. Gemini 3.8 Flash-Lite TTS is ranked sixth at 1,235 with a 16-point interval over 2,000 samples. Run the intervals and the finding is unambiguous: Flash TTS spans 1,243 to 1,277, Flash-Lite TTS spans 1,219 to 1,251, and those ranges overlap between 1,243 and 1,251. Twenty-five Elo points look like a gap in a table and are indistinguishable from noise on a blind listening test with these sample counts.

Now put the price beside it.

• List price — Flash TTS is $9.00 per million audio tokens through December 31, 2026, then $18.00; Flash-Lite TTS is $6.00, then $12.00. Text input is $0.50 per million on both. The tier gap is 33 percent of the audio bill, in both windows.

• Normalised rate — Artificial Analysis converts both to a per-million-characters basis so that models billing in different units can be ranked together, and gives $16.5 for Flash TTS and $11.0 for Flash-Lite TTS. Same 33 percent gap, measured a different way.

• Batch and Priority — Batch and Flex run $0.25 in and $4.50 out on Flash TTS against $0.25 and $3.00 on Flash-Lite TTS; Priority runs $0.90 and $16.20 against $0.90 and $10.80.

• What actually differs — 130 languages against 101, and the full creative positioning against the high-volume one. Google's supported-models table lists single-speaker synthesis, multi-speaker synthesis, voice design and voice replication as available on both.

Read as a whole, that is a weaker case for the premium tier than the marketing implies, and it is a better answer than the umbrella name deserves. Pay the 33 percent if your script spans a language outside the 101, or if you have listened to both and prefer one. Do not pay it because a table shows a higher number, because the numbers overlap.

The third name that keeps getting folded in

There is a second collision underneath this one, and it produces more bad advice than the first. Google shipped two audio products eight days apart in September, and they are not the same kind of thing at all.

Gemini 3.8 Flash TTS is a text-to-speech model: text in, audio out, on the Gemini API's TTS endpoints, built for "scenarios that require exact text recitation with fine-grained control over style and sound." Gemini 3.8 Live is a speech-to-speech system on the Live API, shipped September 15, 2026, built for interactive unstructured audio with multimodal input and tool calls running in the background. Google's own documentation draws the line explicitly, and it is the right line: one reads a script, the other holds a conversation.

The two get merged anyway, in summaries that describe "Gemini 3.8 TTS" as a voice-agent product, which then leads people to evaluate a narration model against a latency budget it was never designed for. If your requirement is a support line that answers the phone and hears the caller interrupt, Flash TTS is the wrong product regardless of tier — and if your requirement is 40,000 words of audiobook with a consistent room tone, the Live API is the wrong one. Deciding which of those you are building is a much larger decision than which tier you pick, and it is the one the naming confusion costs you.

A third string, gemini-3.8-flash, appears on the TTS documentation page too, as the text model used to generate transcripts. It is not a speech model. That one is at least easy to catch.

Choosing, and keeping the choice cheap

The tier decision itself is a string. Both endpoints share a schema, share the voice library, share voice design, and answer on the same speech endpoint with the same authentication, so switching is a one-line change and a re-run — which means the cheapest way to settle it is to generate the same script twice and listen, rather than to argue from a leaderboard at all.

What makes that cheap in practice is not treating the model name as an architectural commitment. On OrcaRouter, google/gemini-3.1-flash-tts-preview — the tier this generation replaces — sits in a catalogue of 200+ models behind a single key, priced at the $1.00 and $20.00 per million tokens Google publishes, with provider list price passed through and no markup added. That pass-through is what makes a tier experiment free rather than a procurement exercise: the same credential you use to compare Flash TTS against Flash-Lite TTS is the one you use for every other model in the catalogue, and automatic failover means a tier you are only trying out never becomes a single point of failure for a production narration queue.

Two things to watch, both of which would change this article. Google lists voice remixing — adjusting timbre, pitch, pace and accent on a voice you already have — as coming soon rather than shipped, and if it lands on one tier first it becomes a real differentiator in a way the Elo gap is not. And if Google ever publishes a non-Flash Gemini 3.8 TTS endpoint with its own identifier and rate card, the umbrella name will finally have something under it, and the comparison will stop being a naming question and start being a genuine tier one.