ACE Music v1 vs Lyria 3: The Leader That Ships Nothing to Developers Against the Model That Ships Only an Endpoint

Google's Lyria 3 arrives in the Gemini API with a price tag and a model identifier. ACE Music v1 arrives with a web app and a leaderboard crown. That inversion — the better-ranked model being the harder one to reach, the lower-ranked model being the one you can call in a curl request — is the whole of this comparison in one sentence, and it is why the usual "which sounds better" framing is close to useless here. ACE Music v1 is ACE Studio's full-song model, ranked #1 on Artificial Analysis's AA-Music-Vocal v1.1 at 1,160. The Lyria 3 family is Google's short-form and full-song generation line, with Lyria 3 Pro sitting at 1,048 on the same board and Lyria 3.5 at 1,050 — seventh and sixth respectively, and both reachable through the Gemini API at $0.08 a song.

The two-point spread between Lyria 3 Pro and Lyria 3.5 is inside both models' confidence intervals, so treat those two as one entry for ranking purposes. The 112-point gap between ACE Music v1 and Lyria 3 Pro is not.

What "Lyria 3" actually names

Google's naming is the first thing to get straight, because three different models share the family name and the pricing page treats them as distinct products.

• Lyria 3 Clip Preview — model ID lyria-3-clip-preview, thirty-second output, $0.04 per song, paid tier only

• Lyria 3 Pro Preview — model ID lyria-3-pro-preview, full-song output, $0.08 per song, paid tier only

• Lyria 3.5 — model ID lyria-3.5, full song, $0.08 per song, paid tier only, released 29 July 2026

Only the two full-song models are on the music leaderboard: Lyria 3 Pro at 1,048 and Lyria 3.5 at 1,050, with 3,191 and 3,203 comparisons respectively. Neither is close to ACE Music v1's 1,160, and unlike the ACE-versus-Suno pairing, the intervals here separate cleanly — ACE's lower bound of 1,139 sits 89 points above Lyria 3.5's point estimate of 1,050. If you trust the board, the vocal result is not a coin flip.

What the board does not do is tell you which of those three Google models to use, or whether the one that scores higher is the one that fits your job. A thirty-second clip at half the price is a genuinely different product from a full song, and the leaderboard does not rank the clip model at all.

The control surface, side by side

Lyria's design assumes you have a specification. ACE Music's assumes you have an idea. That difference runs through every input both models accept.

• Starting point — ACE Music v1 takes text, lyrics, a hummed recording, notes played on a keyboard, MIDI, an image, or two reference songs for a mashup; Lyria 3 takes a text prompt with structured controls for genre, mood, BPM, key and structure

• Melody control — ACE Music builds around a melody you supply and can export the resulting score or MIDI on eligible plans; Lyria 3 Pro works from prompt-level description and offers section-level regeneration rather than note-level input

• Length — ACE Music generates complete songs without a published cap; Lyria 3 Clip is fixed at thirty seconds, Lyria 3 Pro and 3.5 produce full songs

• Watermarking — every Lyria output carries Google's SynthID; ACE Music publishes no equivalent provenance marking

• Editing — ACE Music hands results to ACE Studio, a DAW with note-level vocal editing, for Artist-tier subscribers; Lyria edits inside Google's own surfaces by repainting a section

• Developer access — Lyria 3 is an API model with per-song pricing; ACE Music v1 has no published endpoint

The melody row is the one that decides most real workflows. If your process starts with a hook you hummed into your phone, ACE Music is built for that and Lyria is not — Google's model will interpret a description of a melody, not reproduce one. If your process starts with a brief — "ninety-second upbeat synth bed, 118 BPM, A minor, no vocals" — Lyria's structured controls are more precise than a sentence of prose, and its tempo and key fields are things ACE Music's prompt box does not expose as parameters.

What each one costs to run at volume

Google prices per song. ACE Studio prices per seat. Converting between them is the exercise that usually settles these decisions, and here it produces a genuine surprise: at low volume Google is cheaper, and at high volume ACE Music is dramatically cheaper.

• Lyria 3 Clip — $0.04 per thirty-second song, so $0.08 per minute of audio

• Lyria 3 Pro and Lyria 3.5 — $0.08 per full song, no free tier

• ACE Music Creator — $7.90/month billed yearly for 1,000 credits, about 500 songs, roughly $0.016 a song

• ACE Music Artist — $16.58/month billed yearly for 2,500 credits, about 1,250 songs, roughly $0.013 a song

• ACE Music Free — 20 credits a day, up to 10 songs, no downloads and no commercial rights

Five songs a month through the Gemini API costs forty cents. The same five through ACE Music costs seven dollars ninety, because you are buying a subscription rather than a meter. Flip the volume and the ordering inverts hard: fifty songs a month is four dollars on Lyria and still seven ninety on ACE; five hundred songs is forty dollars on Lyria and still seven ninety on ACE, with downloads and commercial rights included on both.

The break-even sits somewhere near a hundred songs a month. Below it, per-song pricing wins on cost and wins decisively on commitment — no subscription, no annual term, no seat. Above it, the subscription model is not close.

One caveat on the Lyria price: $0.08 is the published rate for a single generation, and generations are not free to iterate on. If your process is nine re-rolls per keeper — which is normal for music — the effective cost per usable track is closer to seventy cents than eight. ACE Music's credit math has the same multiplier but a much smaller base, which is why the subscription model survives heavy iteration and the metered one does not.

Provenance is Google's quiet advantage

Every Lyria 3 output is watermarked with SynthID, Google's imperceptible provenance signal, and that watermark is designed to survive re-encoding. For a brand, an agency, or anyone who has to answer the question "was this made with AI" to a distributor or a platform, that is a feature with real operational value: it makes the answer demonstrable rather than a claim.

ACE Music publishes nothing equivalent. Its commercial terms grant usage rights on paid plans, and its outputs are ordinary audio files with no embedded signal that we can find documented. That is not a defect — most music models ship this way — but it is a difference, and it points the same direction as everything else in this matchup: Google built for institutional buyers who need an audit trail, ACE Studio built for individual creators who need a good song.

The licensing posture differs too, though both are murkier than a buyer would like. Google has said Lyria is trained on licensed material and faces litigation over earlier Lyria generations trained on YouTube recordings, which it is contesting. ACE Studio has published no training-data disclosure that we can locate. Neither company gives you the indemnity a broadcaster would ask for at the self-serve tier.

The access question, and why it cuts both ways

For a developer, Lyria 3 is the only one of these two that exists. A model identifier, a documented request, a per-song price, and an output you can pipe into your own storage. That is the entire integration story, and it is why Lyria shows up in so many prototypes despite a leaderboard position in the middle of the pack.

ACE Music v1 is the mirror image: the best vocal score on the board and no way to call it. No published API, no inference endpoint in its documentation, no developer platform. Its product is a browser — a very good browser, with MIDI input, stem separation, a mashup tool and a bundled DAW — but a browser nonetheless.

That asymmetry has a practical consequence worth naming. Teams that need both — a pipeline that generates at scale and a human who shapes the good ones — end up running two integrations with two vendors and two billing relationships. The alternative is to treat the generation step as a routing problem rather than a vendor choice: one endpoint, many models behind it, failover when a provider degrades, and a model string that changes without a code change. OrcaRouter's version of that is 200-plus models behind a single API at 0% markup on provider list price, with a routing DSL for composing calls and automatic failover across providers. Music models are not in that catalogue today, so this is a description of the architecture rather than a claim about availability — and it is still the right shape for a category where the best model and the callable model keep being different companies.

Where each one wins

ACE Music v1 wins on singing, on melody-first workflows, on price at volume, and on the editing handoff — a generated song that opens in a real DAW is a capability Google's offering does not match at any tier. It also wins on instrument coverage of the process: humming, playing, MIDI upload and mashups are four different front doors into the same model, and Lyria has one.

Lyria 3 wins on access, on commitment, on short-form economics, and on provenance. If you need a thirty-second sting for a social cut, $0.04 and an API call beats a $7.90 subscription and a browser tab. If you need to prove to someone that the audio is synthetic, SynthID is the only answer in this comparison.

If you are choosing today and you write songs, the ranking and the price both point at ACE Music. If you are choosing today and you build software, the ranking is irrelevant until there is an endpoint, and Lyria 3 Pro at eight cents is the pragmatic call — with the caveat that you are buying the seventh-best vocal model on the board and you should not expect it to surprise you.

The thing to watch is not a new version number from either company. It is whether ACE Studio publishes an API. That single decision would collapse the most interesting difference in this matchup, and it is the one change that would make the two-point gap between Lyria 3 Pro and Lyria 3.5 stop mattering.