Lyria 3.5 in Google Flow Music: Sharper Vocals, Smarter Lyrics, and a 3-Minute Ceiling — With No Benchmarks to Check

On July 29, 2026, Google DeepMind launched Lyria 3.5 inside Google Flow Music — the third Lyria model Google has shipped in five months, and the first to arrive inside a dedicated music product rather than the Gemini app. The pitch is four-fold: richer musicality, better lyrics, more expressive vocals, and direct control over tempo and duration. What's more notable is what wasn't included — no benchmark scores, no head-to-head against Suno or Udio, no training-data disclosure, and, at least in the launch post, no mention of SynthID.

That combination is the story. Google is shipping a major music-model upgrade into a free product with no numbers attached, at a moment when the two independent leaders in the space, Suno and Udio, are both fighting Sony Music in court. Here's what Lyria 3.5 actually claims, what it costs, the API question nobody has answered, and how to read the launch honestly.

What Google says Lyria 3.5 actually does

The launch post is short on specifics and long on adjectives. Every claim below is Google's own, with nothing independently measured yet. The four advertised advances: Musicality — "richer, more complex melodic structures that sound more natural." Lyrics — "higher quality lyrics with improved prompt adherence and structural awareness." Vocals — "more realistic and emotionally nuanced vocals, plus improved pronunciation." Creative control — "more easily control the tempo and duration of your outputs."

The DeepMind Lyria model page fills in the product picture that the post leaves out: tracks up to three minutes long, image-to-music (upload an image and ask for a track), vocals in multiple languages — the page's demo prompts span English, Spanish, French, Hindi, and Portuguese — and granular controls over vocal style, acoustic preference, tempo, BPM, key, and instrumentation. There's also a headline editing feature called Selective Section Painting: you can rework a specific part of an existing track, or extend a short melody into a full song without starting over.

Where Lyria 3.5 lives: Google Flow Music

Lyria 3.5 is not a standalone app or a developer endpoint; it ships inside Flow Music, a product with a winding history. It began as Riffusion, the viral open-source experiment of 2022, launched commercially as ProducerAI in 2025, was acquired by Google in February 2026, and rebranded Flow Music in April. Today it is Google's main song-creation surface: conversational editing, section-level reworks, stem separation, and music-video generation built on Veo. Lyria 3.5 is a free upgrade for anyone already in Flow Music, and a free starting point for anyone new — no separate subscription, no waitlist.

What Lyria costs

Flow Music runs on a free tier plus three paid tiers. The free tier grants 500 credits on signup; paid tiers are Starter at $6 per month, Plus at $18 per month, and Member at $48 per month, with the higher tiers adding more credits and more simultaneous generations. Google AI Plus, AI Pro, and AI Ultra subscribers get the corresponding Flow Music tier folded in.

The developer-facing number is separate. For the Lyria family, Google prices a 30-second clip at roughly $0.04 and a full three-minute song at roughly $0.08 — vendor-reported figures for the Lyria 3 / Lyria 3 Pro generation, not yet confirmed for 3.5. At that per-track price, Google undercuts virtually every competitor on raw generation cost. It's also the number that will matter the most if a Lyria 3.5 endpoint ever appears, because it would reset what a per-song API call should cost.

The API question nobody has answered

Lyria 3.5 is confirmed in Flow Music. Whether it reaches the Gemini API or Google AI Studio for developers is not. Google has not said whether Lyria 3.5 replaces Lyria 3 Pro on the developer platform, and the DeepMind model page lists no model ID or pricing contract for it — just the product page and an AI Studio link for the broader Lyria family. Historically Lyria has been deliberately consumer-first: the API surface (the lyria-002 model on Google Cloud, the experimental Lyria RealTime on the Gemini API) has lagged the consumer launches each time. As of today, consumers get Lyria 3.5; developers get an unspecified future.

The field Lyria 3.5 is stepping into

Google's upgrade lands against the two companies that defined this market. Suno has roughly 2 million paid subscribers and about $300 million in annual recurring revenue (figures its CEO reported in February 2026), its v5.5 release added voice cloning and a built-in DAW, and its tiers run $10 and $30 per month — with a track-length ceiling around eight minutes on the top plan. Udio settled with Universal, Warner, Merlin, and Kobalt, but in exchange disabled audio downloads while it rebuilds as a licensed "walled garden," which has frustrated exactly the producers it was built for. And both companies are still in litigation with Sony Music; Suno lost a first-instance ruling in Munich on July 31, 2026.

Google, by contrast, is the only major player in the trio not currently facing a label lawsuit — though it has its own pending case. Independent musicians and songwriters sued Google in March 2026 over Lyria 3, alleging training on copyrighted recordings pulled from YouTube — the complaint claims "at least 44 million clips and 280,000 hours of music." Google moved to dismiss in June, arguing YouTube's terms grant it a broad license to the uploaded material. Google says it trains on data it has rights to, but has never specified what that means for Lyria. That open case is the reason the "licensed data" line in every Lyria article carries an asterisk.

The honest catch list

• No benchmarks. Google published no scores, no listener-preference data, no head-to-heads. The four upgrade claims are qualitative; the honest test is "judged by ears, not metrics."

• No API means no independent benchmarking. Lyria 3.5 is only reachable through Flow Music, so outside researchers cannot run it against Suno v5.5 or Udio under controlled conditions. The only independent testing so far is manual.

• A three-minute ceiling. Full songs cap at three minutes, versus roughly eight minutes on Suno's top plan. Fine for pop songs and hooks; a real constraint for anything longer.

• Training data still undisclosed. The launch post dropped the copyright-framing language that accompanied the Lyria 3 launch — no "designed for original expression" phrasing, no named authors — and the musicians' lawsuit remains pending.

• SynthID is real but unadvertised here. The launch post never mentions watermarking, but the DeepMind model page states every track is imperceptibly watermarked with SynthID. Verifiable provenance is there; Google just chose not to lead with it.

• Early independent impressions are mixed. A community test on Linux.do compared Lyria 3.5 against Lyria 3.0 Pro and found vocals and melody clearly improved — but Chinese lyric recognition severely regressed, to roughly 5% accuracy on characters over five strokes. A hands-on test at Gigazine found Japanese vocals convincing and image-to-music impressive, alongside occasional pronunciation quirks and a video download that errored repeatedly. Directionally better music, weaker non-Latin-script lyrics, is the honest summary of the evidence that exists so far.

What it means for builders

If you're building a music product, the realistic architecture in 2026 is a generation model paired with a text brain that writes the lyrics and steers the prompt — and that text brain is exactly what a router fronts. OrcaRouter gives you 200+ models behind one OpenAI-compatible endpoint at provider list price, passed through with zero markup, so when a vendor cuts a price it's live on our side the same day; automatic failover is the safe way to put a just-released, unproven model on a real path beside a proven one. The moment Google confirms a Lyria 3.5 endpoint — or any music generator reaches an API you can actually call at scale — that's the pattern you'll want: route to the new model, fall back to a stable one if latency, cost, or quality disappoints. We don't route Lyria itself, and we won't pretend otherwise; the routing play here is the layer around it, not the music model.

What to watch next

• A Lyria 3.5 model ID on the Gemini API or AI Studio. That's the confirmation the upgrade has a developer surface, and it will settle the replace-or-sit-beside question for Lyria 3 Pro.

• The first controlled head-to-head. Someone running identical prompts across Lyria 3.5, Suno v5.5, and Udio with native-speaker scoring would be the first external evidence the vocals claim is real.

• Whether the 3-minute ceiling moves. Suno and Udio both went longer as they matured; Google's video-music integration with Veo gives it a distribution lever neither rival has.

• The musicians' lawsuit. If the motion to dismiss fails, Lyria 3.5's "licensed data" positioning gets much harder to sustain — and that would matter more than any vocal upgrade.

FAQ

Is Lyria 3.5 free?

Yes. It ships inside Google Flow Music, which has a free tier that grants 500 credits on signup. Paid tiers run $6, $18, and $48 per month for more credits and more simultaneous generations.

Can developers call Lyria 3.5 through an API?

Not yet confirmed. Lyria 3.5 is available in Flow Music, but Google has not announced a Lyria 3.5 model ID, API pricing, or whether it replaces Lyria 3 Pro on the developer platform. Historically Lyria's API surface has lagged its consumer launches.

How does Lyria 3.5 compare to Suno?

Different products more than different models. Suno v5.5 is an established standalone generator with voice cloning, a built-in DAW, roughly eight-minute tracks, and around 2 million paying subscribers — but it's still in litigation with Sony Music. Lyria 3.5 is a free upgrade inside Google's product, capped at three minutes, with no published benchmark to compare. Google's vocal and lyric claims are unverified; Suno's position is established by heavy real-world use.

Is Lyria 3.5 trained on licensed data?

Google says it trains on material it and YouTube have rights to use, but it has never disclosed the details. Independent musicians sued Google in March 2026 over Lyria 3, alleging training on copyrighted YouTube recordings, and Google moved to dismiss the case in June. The case is still pending.

Lyria 3.5 is a real launch, free to try, and Google's fastest product cadence yet in music — but it is also a launch with no numbers, a three-minute ceiling, and an unanswered API question. Treat the upgrade claims as directional, listen with your own ears, and keep one eye on the Gemini API docs: the moment a Lyria 3.5 ID appears there, this stops being a consumer story and becomes a developer story.