Google's Lyria 3.5 writes full songs for 8 cents apiece

Google DeepMind's newest music generator can now turn a text prompt into a complete song, verses, chorus, bridge, vocals and instrumentation included, and it costs roughly eight cents to do it through the Gemini API. Lyria 3.5 became broadly available on September 4, 2026 across Google AI Studio, the Gemini API, and the Gemini app.

The model actually debuted earlier, arriving inside Google's Flow Music platform on July 29, 2026, weeks before its wider API rollout. Within Flow Music, generation is included at no extra cost, while developers pulling the model through the Gemini API pay $0.08 per song. As of September 3, 2026, Google's own Gemini API changelog still listed Lyria 3.5 as a public-preview model, a status that typically precedes broader stability guarantees and pricing changes.

What it generates and how

Songs produced by Lyria 3.5 run anywhere from 30 seconds to a full three minutes, giving developers enough range to cover jingles, background scores, or radio-length tracks. The model responds in whatever language the prompt is written in, and Google specifically notes support for French alongside English.

A notable input feature lets users feed up to ten images into the model to steer the mood, genre, and instrumentation of the output, effectively letting visual references double as a style guide. Under the hood, Lyria 3.5 relies on latent diffusion, the same broad technique that underlies most modern image generators, applied here to audio waveforms instead of pixels.

Pricing shift and provenance

The release also reshuffles Google's music-generation lineup. The prior model, now rebranded Lyria 3 Clip, has been pushed into legacy status and priced at $0.04 for a 30-second clip, roughly half the per-second cost of the new flagship once duration is factored in.

Every clip Lyria 3.5 produces carries an inaudible SynthID watermark, Google's audio fingerprinting system meant to let platforms and researchers trace AI-generated tracks back to their source. Google also states that its music models are trained exclusively on licensed data, a detail likely aimed at preempting the copyright disputes that have dogged rival generative-audio tools.

For teams building music or audio features into products, the combination of per-song pricing, image-conditioned style control, and multi-minute output length makes Lyria 3.5 one of the more developer-oriented entries in Google's generative media stack, sitting alongside Veo for video and Imagen for stills.