Gemini 3.8 TTS: Two Pelicans, Two Models, and a Price That Doubles in January

The fastest way to understand what the vendor shipped on September 23, 2026 is to look at the demo that went around with it: two pelicans arguing about whether to move to Pacifica Pier, one voice per bird, scripted turn by turn. Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS are the two models under that demo, both generally available on the Gemini API, and the pelicans are not a gimmick. They are the first thing in this generation that the previous Gemini speech models could not do at all: two speakers, one generation, a script that controls who says what.

The price is the other half of the story, and it is where the two models split. Flash TTS bills at $0.50 per million text tokens in and $9.00 per million audio tokens out through December 31, 2026, then $18.00; Flash-Lite TTS bills at $0.50 and $6.00 on the same schedule, then $12.00. Those are Google's own rates. Converted by Artificial Analysis to a per-million-characters basis so they can sit on the same board as everyone else, they come out at $16.50 and $11.00 — roughly a third cheaper for the Lite model, and both well under what the top of that board charges.

So the interesting question is not whether the models are good. It is which of the two you should build on, because the gap between them is not the gap you would guess from the names.

What the September 23 announcement actually contains

Two models, not one, and Google is explicit that they are a split rather than a small and a large version of the same thing. The announcement names five places they land — Google AI Studio, the Gemini API, Gemini Enterprise, Gemini Notebook and Google Vids — and the API identifiers are plain: gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts.

The capability list is longer than the headline suggests. From Google's announcement and the Gemini API speech-generation docs:

• Voice design — you describe a voice in words and get a custom voice ID back, rather than picking from a fixed list.

• Voice replication — a 30-second sample produces a voice ID, which is the path most teams will actually use for a brand voice or a narrator.

• Two-speaker dialogue — the pelican case. The script carries speaker turns instead of a single block of prose.

• Turn-level styling — the docs describe a speech_metadata annotation that attaches direction to a specific turn rather than to the whole request.

• Prebuilt voices, plus an Extended Voice Library reachable at GET /v1beta/voices.

• Three audio encodings — WAV by default, with mulaw and alaw available for telephony-shaped pipelines.

• Text-only in, audio-only out. The docs are blunt about this: TTS models take no other input modality and produce no other output, and there is a separate Limitations section listing the restrictions.

What is not in the announcement is any of the numbers a capacity planner needs. Rate limits, quotas and the storage lifetime of a designed voice all live in the docs and the pricing page, not in the launch post, which is the usual reason a launch week goes smoothly for demos and badly for production.

The pelicans are the actual news

Single-speaker TTS has been a solved-enough problem for two years. What was missing was a model that would hold two identities apart across a long script without drifting, and that is what the demo is really testing — not whether the voice sounds human, but whether speaker A is still speaker A four minutes in.

Two things follow from that, and they are the reason this matters beyond podcast toys.

The first is that the unit of work changes. With a single-speaker model, a twenty-minute dialogue is twenty minutes of audio that you stitch from two independent runs, and every seam is a place where prosody resets. With a two-speaker model, it is one call, one context, and the model can react to the line before it. That is a quality difference you cannot buy back with post-processing.

The second is that the script format becomes the interface. If your pipeline already emits something SSML-shaped — an annotation per phrase — this generation is close to a drop-in. If your pipeline hands a model a wall of text and hopes, you now have a reason to restructure it, because the styling that used to be implicit is now something you have to say out loud. That is the same trade-off the rest of the field is making in different ways, and it is the axis on which these two Google models compete with everything else on the board.

What an hour of two-pelican audio costs

The token maths is worth doing once, because the per-million-audio-token figure is not a per-hour figure and the difference has surprised people.

Audio tokens are produced at a fixed rate per second of output, so cost scales with the length of the finished audio, not the length of the script. Take Google's own rate for Flash TTS — $9.00 per million audio tokens out — and the arithmetic for a one-hour finished piece lands in the single-digit dollars on the standard tier, with the Lite model at two-thirds of that. Batch and Flex pricing, at $0.25 in and $4.50 out for Flash TTS against $0.25 and $3.00 for Flash-Lite TTS, cuts it further for anything that does not need to be generated on demand. Priority, at $0.90 and $16.00, is for the work that does.

Three practical consequences:

• Regenerating a paragraph is cheap; regenerating a series is a decision. At these rates the expensive part of a speech pipeline stops being inference and goes back to being the human who approves the take.

• The text-input rate is noise. $0.50 per million text tokens on a script of a few thousand words is a rounding error next to the audio side. Do not optimise the script for length.

• The December 31 cliff is real and it doubles the audio rate. If you are planning a 2027 rollout, budget the post-promotional number, not the launch number, or you will be re-planning in January.

The number that is independent, and the ones that are not

One figure in this launch comes from somewhere other than Google. On the Artificial Analysis Provider Voice Arena, captured September 24, 2026, Gemini 3.8 Flash TTS sits at rank 2 with an Elo of 1,260 across 1,999 samples and 8 arena voices, and Gemini 3.8 Flash-Lite TTS sits at rank 6 with 1,235 across 2,000 samples. Both are inside each other's confidence intervals at ±16 and ±17, so the board is telling you they are statistically indistinguishable on preference — which is a genuinely useful result, because it means the cheaper one is not measurably worse to listeners.

Everything else is Google's own claim and should be read as such: the expressiveness language, the language counts, the replication quality. The arena gives you a preference ranking and nothing else. It does not tell you how either model handles a 40-minute script without drift, how it behaves on a proper noun it has never seen, or what happens to a designed voice after a week.

The Lite model is also the better argument for the pair being a real split rather than a marketing tier. A 25-point Elo gap that sits inside the error bars, against a 33% price gap, is the shape of a decision that price-sensitive pipelines should make in favour of Lite almost every time — and the shape of a decision that latency-sensitive pipelines should make in the other direction, since the two differ on the clock as well as the bill.

Where to run it, and the honest version of that answer

OrcaRouter does not host the Gemini 3.8 TTS endpoints. That is worth stating plainly rather than implying otherwise, because the useful thing we can say about these two models is not that we serve them — we do not — but that a vendor price cut like the December 31 change is exactly the kind of event that makes routing worth having.

OrcaRouter puts 200-plus models behind a single API key at provider list price with no markup, so a vendor's rate card is what you pay, and a change to it is live on our side the same day rather than at the next billing cycle. For a speech pipeline that is mid-migration, the failover layer matters more than the catalogue: you can put a request path on a model that is still being evaluated without betting a production route on it, because a failing call can fall through to something that is already proven. The routing DSL composes several models into one call for the cases where no single voice is good enough, and model fusion answers a prompt with a panel when the answer matters more than the latency.

For the Gemini 3.8 TTS models themselves, the place to call them is Google's own API, and the surfaces Google named on September 23. What the pair does to your architecture — one call for two speakers, styling as an annotation, a voice that can be described rather than chosen — is the part that outlives the launch discount.

What to watch next

Three things will decide whether this generation is a step change or a feature.

The first is drift. Nobody has published a long-form evaluation of whether the two-speaker path holds identity over a full hour, and until someone does, the pelican demo is a demo. The second is voice lifetime. A designed voice that expires in a week is a prototype; a voice that can be re-derived from a stored config is a product, and the answer depends on which of the two identifiers you keep. The third is what happens after December 31, when the promotional rate ends and the per-hour cost of a speech pipeline doubles overnight.

None of those are visible from the launch post. All three are visible from the docs, which is where the launch coverage stops reading.