ElevenLabs releases SFX v3 API for multi-part sound effect generation

Details

ElevenLabs released SFX v3, a sound effects model that follows multi-part instructions more closely than SFX v2 and handles vocals and music better.
Users can set model_id to eleven_text_to_sound_v3 on the POST /v1/sound-generation endpoint to access the new model.
The model automatically determines sound length based on the prompt, though users can specify duration_seconds for exact timing.
Raising prompt_influence above the default 0.3 makes the output follow the prompt more closely and reduces variation between generations.
Looping functionality is currently only available for eleven_text_to_sound_v2, requiring users to stick to v2 if they need loop: true.
Prompting best practices include listing events in chronological order and using concrete, specific wording to improve results.
SFX v3 is less likely to clip on high-frequency sounds like gongs and chimes, and the engineering team reports it sounds more natural overall.
Both models run on the same endpoint, allowing for cheap A/B testing before switching from v2 to v3.

Key point