ElevenLabs Ships Music v2.5 With Licensed-Only Training Data
ElevenLabs released Music v2.5, an upgraded AI music generator that beat its predecessor in a blind listening test of nearly 48,000 pairs.
ElevenLabs has launched Music v2.5, the latest version of its AI music generation model, and is making it available through both its consumer app and its developer API. The company says the upgrade was trained exclusively on licensed music, a detail that separates it from many rivals facing copyright litigation. The release arrives as the music industry continues to press AI developers on where their training data comes from.
The headline number behind the launch is a preference test: in a blind study covering close to 48,000 comparison pairs, listeners chose Music v2.5 over the previous generation. ElevenLabs did not publish the full methodology, so it remains unclear how the pairs were selected, who the listeners were, or whether the results hold across genres. Still, the scale of the test is notable for a sector where qualitative claims often outrun measurable evidence.
The access model follows ElevenLabs' existing playbook. Music v2.5 is available on a free tier, with paid pro options that likely unlock higher usage limits or commercial rights, mirroring how the company distributes its voice products. Putting the model in the API matters as much as the app release: it lets third-party developers build music generation into their own tools rather than sending users to ElevenLabs' interface.
Licensed-only training is the claim that will draw the most scrutiny. Several AI music companies have been sued by major labels over unlicensed use of recordings, and the outcomes of those cases could reshape what training data is legally viable. If ElevenLabs can document its licensing chain, that becomes a selling point for enterprise customers wary of legal exposure. The company has not yet detailed which rights holders it signed with or what those deals cover.
The competitive field is crowded. Suno and Udio have built large user bases, while Google and Meta have pushed their own audio generation research, and Stability AI has experimented in the space. ElevenLabs' advantage has been distribution: a large existing base of voice and audio users who can be moved onto a new music product with little friction. Whether that converts into sustained music usage is a different question from whether the model scores better in a listening test.
What to watch next is documentation. ElevenLabs says the training data is licensed, but the proof will come in the form of named partners, licensing terms, and whether rights holders confirm the arrangements publicly. A preference win in a blind test is a marketing asset; a verifiable licensing trail is a legal one. The company has not said when, or if, it will publish either.