Google put Lyria 3.5 into the Gemini API on 4 September. The announcement calls it "our best-sounding music generation model" and gives it "more expressive vocals and richer musical arrangements". It gives no price.
The price is on the developer pricing page, stamped the same day, and it is more interesting than the announcement.
Lyria 3.5 has exactly one rate: $0.08 per song. That is the same figure Google charges for Lyria 3 Pro Preview, the full-song model it supersedes. A better model turned up and the rate card did not change.
The other half of the story is the SKU that did not turn up with it. The only 30 second rate Google publishes is $0.04, it belongs to lyria-3-clip-preview, and the pricing page files that model under a description it gives nothing else in the section: "Google's family of legacy music generation models".
What the pricing page lists#
Three rows, two prices, and no free tier on any of them. If you want to hear what Lyria sounds like through the API, the cheapest way to find out is $0.04.
The full song is exactly twice the clip. That is the only ratio in the section, and it is worth holding on to, because the clip is the row that is not on the new model.
The price that did not move#
Lyria 3.5 (Full Song) is $0.08. Lyria 3 Pro Preview (Full Song) is $0.08.
Google describes Lyria 3.5 as delivering higher fidelity than what came before, and charges the same for it. That is a real pattern in model pricing this year rather than a Google quirk: the list price holds and the quality underneath it moves, which is the same shape as a model getting cheaper without a price cut. It is also the opposite of what Gemini Flash pricing did, where the number itself carried a date.
Flat pricing across a version bump is good news for anyone who has already budgeted. It is worth noticing anyway, because it means the upgrade decision has no cost side. There is nothing to trade off.
The clip you can still buy is filed as legacy#
Google's two pages describe the Lyria 3 family differently, and both were updated on 4 September.
The music generation docs open by saying "Lyria 3.5 is Google's family of music generation models", then list two members:
The pricing page puts the same clip model under "Google's family of legacy music generation models".
So the only way to buy 30 seconds of Lyria is a preview-tagged model that one Google page presents as current and another files as legacy. Both descriptions are Google's, published on the same day, about the same model ID.
Google has not announced a deprecation date for either Lyria 3 model, and nothing on either page says one is coming, so this is a note about the shape of the catalogue and not a prediction. If your product depends on 30 second clips, it depends on lyria-3-clip-preview specifically, and that is worth writing down somewhere you will read again.
One more thing the announcement says that the docs do not: in the Gemini app you get "the flexibility to choose short or longer tracks". That is the app. Through the API, short means the clip model.
What a song costs next to a video#
The same pricing page sells video. Veo 3.1 Lite, the cheapest video rate Google lists, is $0.05 per second at 720p.
Google says a Lyria 3.5 song runs "a couple of minutes". Take Google at its word and call that two minutes. Two minutes of the cheapest video Google sells is 120 seconds at $0.05, which is $6.00. Two minutes of the music model is $0.08.
Video costs 75 times what music costs for the same length of output.
That single ratio explains a lot about which generated media is turning up inside shipping products this year and which is still a demo.
The two limits that change the real price#
The docs pin down two behaviours that matter more to a budget than the rate card does.
Generation is single turn. In Google's words, "Iterative editing or refining a generated clip through multiple prompts is not supported in the current version of Lyria 3.5." You cannot nudge a track towards what you wanted. You prompt again.
And results are not deterministic: "Results may vary between calls, even with the same prompt."
Put those together and $0.08 is the price of a generation, not the price of a song you keep. Every rejected take is another $0.08.
Google publishes no success rate and this piece does not have one either, so use your own. At one usable track in three, a usable song costs $0.24. At one in five it is $0.40. The rate card is the easy half of the estimate; the retry multiplier is the half you have to measure yourself, which is the same discipline as costing a model in production.
The rest of what is published#
Both models produce 44.1 kHz stereo MP3 and both accept text or image prompts through Google's Interactions API. Google describes the output as delivering "structural coherence, including vocals, timed lyrics, and full instrumental arrangements".
On the paid tier, the pricing table's last row says your input is not used to improve Google's products. Since no Lyria SKU has a free tier, that is the only row that can apply to an API caller.
Lyria RealTime, Google's streaming music model, is linked from the pricing page navigation and has no rate row on it. A rate may be published elsewhere in Google's documentation. It is not on the page that prices Lyria 3.5.
What to take from this#
Three things, in the order they will cost you money.
The upgrade is free, so take it: $0.08 buys a better model than it did last week.
Short clips are a different model with a different price and a legacy label, so name lyria-3-clip-preview explicitly in anything you build rather than assuming the family will carry it forward.
And the number on the rate card is a per-attempt number. A model you cannot iterate on and that answers differently every time is priced per roll of the dice, which is a structure worth recognising anywhere it appears, including in a model with one price and no discounts and in a price cliff you only meet at scale.
Discussion