On August 25, 2026, ElevenLabs introduced Composer, a song editor for its AI music platform ElevenMusic. The tool is built on the Music v2 model and changes the workflow for AI music: after generation, a track can be opened section by section and specific fragments — a verse, chorus, or outro — can be refined without regenerating the entire composition.

image
image
image

What happened

The announcement took place on August 25, 2026: ElevenLabs published an announcement on the company's official X account and a demo video on YouTube. In Composer, a generated track opens as a document divided into sections (verse, chorus, bridge, outro), and each of them can be edited independently: rewrite the text of only one verse, change the duration or style of a selected fragment, rearrange parts, duplicate a chorus, insert a new outro. When locally regenerating one section, the other blocks of the track are preserved, and several versions of one fragment are generated and displayed side by side to choose the best one. The first step is to choose a starting point: an empty project, a text prompt, your own lyrics, or a ready-made track.

Context

The main novelty of Composer is not in the model: Music v2 remains the same ElevenMusic engine, the company added a management layer to it. Previously, the typical generative music scenario looked like "one button — finished track", and if the result was unsuccessful, it remained to regenerate the song entirely — successful fragments were broken in the process, and track refinement was a lottery. The stated capability of Composer — regenerating one section while preserving the rest of the parts — is essentially conditional generation with a fixed context of neighboring blocks, that is, inpainting at the level of song structure. Similar logic has long been familiar in DAWs, digital audio workstations, where the musician themselves sets the track structure and iteratively refines individual blocks.

Why this matters for the industry

For the industry, the announcement is a signal that competition in AI music is shifting from the question of "who will generate a song" to the question of "who will allow it to be iteratively edited". One-time generation with one button risks becoming a mass standard, and value shifts to an iterative workflow: for developers, Composer sets a ready-made template of "generate — refine a section — accept", which competing platforms will probably have to replicate. There are no changes for integration engineers: the tool is intended for end users, and the announcement does not include an API.

Why this matters for users

For the reader, the main effect becomes moving away from the "lottery": if the model only messed up the chorus, there is no longer a need to regenerate the entire track and hope for luck — you can open the song section by section and rewrite a specific place. For video creators, this is a practical tool: separately refine the intro or adjust the finale to the edit without touching the rest of the composition. The barrier to entry is low: according to the announcement and the app description in the App Store, the free ElevenMusic plan provides up to 5 track generations per day, the Pro subscription costs $9.99 per month and opens up to 400 tracks, so you can try the editor right now, without paying for a subscription.

What is still unknown / limitations

There is no evidence of scientific novelty in the sources: Composer works on the already existing Music v2 model, and the announcement does not contain a description of the architecture, training methods, and quantitative metrics. How exactly the preservation of neighboring sections during local regeneration is implemented — a condition on hidden representations, fine-tuning for the editing task, or splicing in post-processing — is not disclosed. Benchmarks are also absent: there is no comparison of section editing with regenerating the entire track, A/B tests, quality metrics, and latency data — in the sources, these are product claims and a demonstration. It is also unclear how section boundaries are determined if a ready-made third-party track is provided as input.

Sources

Author

Look at AI, editorial team