Google's Lyria 3.5 costs four times what a Suno song costs — and it is still the only frontier music model a developer can actually buy

What shipped

Lyria 3.5 is not a new model as of this week — it first appeared in Google Flow Music on 29 July 2026. What changed on 4 September is where you can reach it. It is now in the Gemini app on web and mobile for all users globally, and it is on the Gemini API, which opened in public preview on 3 September.

The developer surface is the substantive part:

Neither has a free tier. Songs are generated in one shot with structure controllable through section tags — Verse, Chorus, Bridge — and optional timestamps. You can pass up to 10 images alongside the text prompt and have the model compose against them. Output is MP3 by default, with WAV available on 3.5. Requests for named artist voices or copyrighted lyrics are refused.

In the consumer app the tiering is conventional: free accounts generate under rate caps with shorter maximum durations; Google AI Pro and Ultra lift the caps and add stem export and faster queues.

The price comparison everyone got backwards

The instinct on a launch like this is to assume the hyperscaler is buying the market. Run the numbers and the opposite is true.

Suno Pro is $10 per month for 2,500 credits. A full song costs 5 credits, so the allocation is about 500 songs — roughly $0.02 per song. Premier is $30 for 10,000 credits, about 2,000 songs, or roughly $0.015. Against Lyria 3.5’s $0.08, Google is charging about four times Pro’s effective rate and more than five times Premier’s.

That comparison deserves one honest caveat, because subscription averages flatter the incumbent. Suno’s per-song figure assumes you actually consume your credits; unused allocation is money spent for nothing. Google’s price is true pay-as-you-go with no commitment and no wastage. So the ranking flips by volume:

• Under ~125 songs/month, Lyria’s pay-as-you-go is cheaper than a $10 Suno Pro subscription. At 40 songs, it is $3.20 against $10.
• Above ~125 songs/month, Suno Pro wins, and the gap widens fast at Premier volumes.

That crossover is the number to run against your own usage. But it does not rescue the “Google undercut Suno” reading, because heavy users — the ones who define the category — sit firmly on the side where Suno is cheaper. Google did not come in under the incumbent. It came in above it, deliberately.

What Google actually took

The thing Google took was not price. It was the developer.

Suno does not sell a public API. Not a restricted one, not an expensive one — none. There is no developer console, no self-service key issuance, no public SDK, no official documentation. Chief product officer Jack Brody said on 1 July that the company was exploring a developer API starting with curated partners, and an intake form went up. Two months later there is still no announced launch date, pricing model or self-serve portal.

What filled that vacuum is a grey market: third-party resellers offering unofficial programmatic access at roughly $0.014 to $0.111 per song. Some are cheaper than Google. None of them is Suno. Building on one means your production dependency is a reseller’s continued ability to reach an interface it was never granted, with no vendor commitment behind it and no recourse when it stops working.

Set against that, $0.08 with documentation, a model ID and a billing relationship is not an expensive price. It is the first quoted price the category has had. This is the same structural point this desk has made about platform layers elsewhere — the Assistants API sunset showed how short the half-life of a vendor’s convenience layer can be — except inverted. There, the risk was building on something that got withdrawn. Here, the risk is building on something that was never offered.

Google has run this play before in audio. When Gemini 3.5 Transcribe reset the speech-to-text price floor in August, the mechanism was the same: take a capability that specialists sold through their own front doors, and make it a line item on an API that developers already have keys for. The bundling advantage is not the Gemini app. It is that the music model now sits behind the same key as Gemini 3.8 Flash, with one bill and one integration.

The part nobody is pricing: the mark

Every Lyria output carries a SynthID watermark. It is applied at the model, on every surface, with no toggle, and Google describes it as imperceptible to the ear.

That matters more this quarter than it would have six months ago. Article 50 of the EU AI Act became applicable on 2 August 2026, and Article 50(2) requires providers of generative systems to mark outputs in a machine-readable format detectable as artificially generated. For a buyer, the useful question is not whether a vendor intends to comply but whether the file you are holding is already marked.

On that test the two vendors are currently in different places. Google’s mark is shipped and unconditional. Suno announced on 6 August — under real legal pressure, following a July German court ruling and with Universal and Sony suits outstanding — that it would add invisible audio watermarking and Musixmatch fingerprinting, with rollout “in the coming weeks.”

Two cautions, because the sharp version of this claim would be wrong:

Suno is not non-compliant. It has committed to exactly what the Article asks for, and a month is not a long delay for an audio watermarking rollout.

The grace period is genuinely ambiguous here. The Act gives synthetic-content systems already on the EU market before 2 August until 2 December 2026. Suno plainly qualifies. Google’s music generation also predates the cutoff — Lyria was in Flow Music from 29 July and in the Gemini app before that — so the neat story where Google had no grace period and Suno did is not supportable. Both have a plausible claim to it.

Which is why the buyer’s test is not legal standing but verifiability. Google’s marking is confirmable today from outside the company. Suno’s rollout status is not something a customer can check. If you publish AI-generated music commercially into the EU, that difference is worth more than the $0.06 per song. It is the same distinction that ran through the Pulitzer disclosure record: a disclosure that exists somewhere is not the same as one that reaches the person who needs it.

And it only covers the provider’s half. Deployer-side duties can still land on you when you publish, which is what the Commission’s Code of Practice on Transparency of AI-generated Content addresses. SynthID removes one compliance question from your build. It does not remove all of them.

What this means for buyers

If you generate music by hand, stay on Suno. It leads on vocal realism, it has ~2M paid subscribers and ~$300M ARR behind its roadmap, and Lyria’s single-turn generation has no answer yet to the regenerate-and-refine workflow heavy users depend on. Nothing this week changes the recommendation on our best AI audio tools list.

If you build products, this is the first real option. A documented API, a published price and a billing relationship beat a cheaper reseller you have no contract with. Run the ~125-song crossover against your projected volume before assuming either side is cheaper.

If you publish commercially into the EU, weigh the mark, not the rate. The delta is $0.06 a song. The cost of establishing marking provenance yourself, after the fact, is not $0.06 a song.

If you are a content creator already paying for Google AI Pro or Ultra, generate in the Gemini app before adding a second subscription. For background music, jingles and ringtones the bundled tier may simply be enough — and the same logic that applies to voice work with ElevenLabs applies here: the specialist is better, and the bundled option is often sufficient.

The bottom line

Google did not launch a price war in AI music. It launched at four times the incumbent’s per-song rate and will likely still take the developer segment, because the incumbent has spent two years not selling to it.

That is the pattern worth carrying forward. The most reliable way into a category with an entrenched consumer leader is not to undercut the leader’s price — it is to sell the surface the leader has left unattended, and to arrive with the compliance property the market has just started requiring. Suno owns the creators. As of 4 September, Google owns the only music API you can put in a production system and point a lawyer at.

Update, 23 September 2026 — Google extended the same marking posture to speech, and added C2PA. Gemini 3.8 Flash TTS and Flash-Lite TTS, released 23 September, carry a SynthID watermark on every generated clip, with C2PA credentials additionally attached to voice-replication output. That is the pattern this article identified with Lyria, now applied across a second modality and strengthened — Article 50(2)‘s machine-readable marking duty is met by the provider before the file reaches you, on every surface, without a toggle. The deployer-side disclosure duties are unchanged. One thing does not carry over: voice replication is unavailable in the EEA, the UK, Switzerland, India, Illinois and Texas, so the C2PA half of that story is not something EU buyers can use. The speech launch, its January price reset and the EU exclusion.

Questions or corrections? Email Pick Right. Want the full list? See all news.