On October 29, ElevenLabs launched SFX v3, a sound-effects model. In a single generation, the new model can present several sounds in sequence, so creators no longer need to split a scene's sound into separate generations and stitch them by hand. The model is live across ElevenCreative surfaces and in ElevenAPI.
One prompt holds a whole stretch of sound
In the older text-to-SFX workflow, a scene with several sounds often had to be generated in pieces: first crowd cheering, then a ball dropping, then the same crowd booing, with timing aligned in editing. SFX v3 accepts prompts containing several instructions and places those sounds in one generation, in the order described. In its official documentation, ElevenLabs gives the example that describing the action, setting, atmosphere, and performance works better than a category word alone — write "rain against a window in a quiet apartment" rather than just "rain."
Beyond generating several sounds at once, SFX v3 also improves handling of music and vocals, with less harshness on high-frequency sounds, so gongs, chimes, and whistles sound cleaner. ElevenLabs says the new model beats SFX v2 across its internal benchmarks, but the full comparison tables have not been published; creators are best advised to run their own usual prompts on both versions before switching.
Where it can be used
SFX v3 covers the main ElevenCreative entry points: Flows, the Sound Effects tab, Studio, and the Music Editor, and it is also available in the mobile app. Developers call it through ElevenAPI with the model ID eleven_text_to_sound_v3; SFX v2 remains the default model, so switching requires changing the model ID.
On specifications, a single effect runs up to 30 seconds, outputs as MP3, and non-looping effects can be output as 48kHz WAV, while long atmospheres can loop seamlessly. Billing follows the existing rules of the sound-effects endpoint: when a duration is specified, billing is per second in credits, and ElevenLabs has not announced any change to per-generation credit pricing.
Three limits remain
First, exact counts are still unreliable: a request such as "hit 4 times" may not be counted correctly, and key rhythms still need checking in post-production. Second, SFX v3 is currently in limited access, and some regions or accounts will hit an unknown-model error when calling it; in that case, the only option is to wait for wider access. Third, it removes the fiddly work of stitching sound effects together, but it does not replace dialogue design, composition, or final mixing — one prompt can sketch a scene's skeleton, and the finished quality still depends on human mixing.
