AI & Creative Tools

ElevenLabs' Music v2 Can Swing From Opera to Metal Mid-Song — and Still Hand You Real Stems

The new model adds inpainting, section-by-section composition, and true stem exports, built entirely on licensed data so generated tracks are commercially clear on release.

AI image and video generation have spent the last year fighting very public battles over training data and licensing. AI music generation has been fighting the same battle more quietly, and ElevenLabs’ new Music v2 model is the clearest sign yet of it being resolved on the audio side, at least for one vendor: trained exclusively on licensed data, with no sync fees or clearance delays before a track can be used commercially.

Watch: How to Create AI Music — Eleven Music tutorial (YouTube)

What actually changed from v1

The headline feature is genre flexibility inside a single generation — a song can move from opera to heavy metal and back without losing musical coherence, which was simply not something the first version of the model could hold together. Underneath that, three more practical additions do more for a working producer: inpainting, which lets you regenerate one section of a track without disturbing the rest instead of rerolling the whole generation; full section-by-section composition, building a real intro/verse/chorus structure rather than a short undifferentiated clip; and meaningfully better handling of complex, fast-paced vocal delivery and lyrics across multiple languages.

The part that matters for actual production work

Music v2 ships across three surfaces — ElevenMusic for hands-on creation, ElevenAPI for developers building it into other products, and ElevenCreative for licensed content at scale — and all three now include true stem downloads: isolated vocals, drums, bass, and instrument layers, not just a mixed master. That closes the gap that made most earlier AI music tools a dead end for anyone doing real post-production — a generated track you can’t remix, EQ per-element, or drop into a DAW alongside other stems is a novelty; one you can is a starting point.

Where this leaves the field

Suno remains the model most people benchmark against, and as of this year there still isn’t an open-source equivalent that consistently matches it — the closest open options land close on some prompts and clearly behind on others. Music v2’s actual contribution isn’t necessarily topping Suno on raw generation quality; it’s making the licensing and stem-export questions boring, which is exactly what needs to happen before AI-generated music stops being a demo and starts being a normal input to a real mixing session.