ElevenLabs launched SFX v3, its new text-to-sound-effects model, on October 29. Its November 2 changelog says the model is available to every workspace. It is now the default in ElevenCreative's Sound Effects, Studio and Flows, and developers can use it through ElevenAPI.
The aim is to handle scenes containing several sounds rather than a single effect. ElevenLabs says the model follows multi-instruction prompts better, improves music and vocals, and reduces harsh audio artifacts. These improvements come from company testing and should be judged against the specific material being produced, not as a guarantee for every attempt.
Several sound sources in one request
Creators can describe the main action and surrounding sounds, then listen to their relationship. A nearby knock and distant rustling, for example, should occupy different places in the sound scene. This is an illustrative use case, not a benchmark: prompts can ask for relationships and sequences rather than simply name an effect.
ElevenLabs also reports better handling of music and singing and more natural sounds such as bells and whistles, where earlier output could become crackly or distorted. That may reduce retries for video and game editors, but the result still needs checking in the final mix, not only as an isolated file.
The launch names Sound Effects, Studio, Flows and Music Editor as creation surfaces. November's update adds that sound effects generated by Video to music also use the new model. The list of interfaces and the list of defaults are not identical claims; each screen need not offer exactly the same options.
API developers also have a migration date
For Create sound effect, specify eleven_text_to_sound_v3 in model_id. The documentation says loop: true creates an effect intended to loop smoothly. Naming the model explicitly helps teams track which version produced files and compare it with existing output before altering production automation.
The eleven_text_to_sound_v2 model is deprecated and scheduled for removal on March 1, 2027. That is a future removal date, not a claim it has already stopped working. Teams should locate relevant calls, prepare listening tests and update the model before the deadline, especially where loop quality or sound character matters.
The update offers a shorter route to trying a complete sound environment. The production test remains whether requested sounds are present, ordered correctly and appropriate in level and texture. Keeping the prompt and model version with the file helps teams revisit a result or reconstruct why they chose it in the next editing round.