THE ESSENTIALS IN 30 SECONDS

  • SFX v3, announced on October 29, is available in ElevenCreative (including the mobile app) and through ElevenAPI under the model ID eleven_text_to_sound_v3.
  • ElevenLabs says it follows prompts describing several sounds more reliably than SFX v2 and handles music and vocals better, based on its own benchmark testing.
  • The documentation says access is limited and may not cover every region; v2 remains the default model.

ElevenLabs, best known for its synthetic voices, has released SFX v3, the third generation of its text-to-sound-effects model. The company announced it on its blog on October 29 and says the model is live both in ElevenCreative, its suite for creators, and in ElevenAPI for developers who want to build sound generation into their own products.

What the company says has improved

The main pitch is scenes with more than one thing happening. According to ElevenLabs, SFX v3 follows prompts that describe several distinct sounds more faithfully than SFX v2, so more of the requested events actually end up in the output. The company also says the model does better with prompts that mix music or singing with ambient noise, and that overall audio is cleaner, with fewer unwanted artifacts on bright sounds such as gongs, chimes and whistles. These claims come from ElevenLabs' own benchmark testing; the blog post gives no figures, and Creators Toolbox notes that the comparison with v2 is qualitative only.

Inside ElevenCreative, the model can be used from the Sound Effects tab, Flows, Studio and the Music Editor, as well as in the ElevenLabs mobile app.

What the documentation adds

The developer documentation lists the new model under the ID eleven_text_to_sound_v3, while eleven_text_to_sound_v2 remains the default. ElevenLabs describes access to v3 as limited and possibly unavailable in some regions; accounts without access receive an "unknown model" error. The docs state that a single generation can last between 0.1 and 30 seconds, with the length chosen automatically when none is set, and that output is available as MP3, plus 48 kHz WAV for non-looping effects. A "prompt influence" setting lets users trade literal interpretation for more creative variation.

Caveats

ElevenLabs itself advises comparing v2 and v3, since results depend on the prompt. The documentation warns that exact repetition counts, such as a set number of hits, are not always respected, and that small wording changes can affect the result. The announcement does not mention pricing specific to v3, and no independent evaluation of the model was available at the time of writing. For full music production, ElevenLabs points users to its separate Music API rather than the sound-effects model.