ElevenLabs SFX v3: what it is and how to use it
2026-10-30 · 4 min read
ElevenLabs SFX v3 is the company's new text-to-sound-effects model. You describe a sound and it generates up to 30 seconds of audio. What's new in v3 is that it does a better job with prompts that ask for several sounds at once, and with prompts that include music or singing. You can use it in the ElevenCreative app, the mobile app and the API. The API docs call it limited access, though, so your account may not have it yet.
@ElevenLabs announced it on October 29 with a short demo. The post opens with "Big scenes rarely come down to one sound," which is the core of the pitch. The same day the company published a launch post with the details.
What changed from SFX v2
A door slam or rain on a roof is one sound, and any model can do that. The harder case is a scene. The launch post's example is a football clip that needs a crowd cheering, a ball dropping, and then the same crowd booing a few seconds later. ElevenLabs says v3 handles prompts like that more reliably than v2, and that it beat v2 on audio quality in the company's own benchmark testing.
The models page lists four improvements over v2:
- It follows several instructions in one prompt.
- It handles prompts that include music or vocals.
- It produces more natural-sounding results.
- It is less likely to clip on high-pitched sounds like gongs, chimes and whistles.
The docs also say results vary by prompt and suggest comparing both models on your own prompts. That's good advice. v2 is still the default model, so test before you switch anything that already works.
How to try it
In the app, the launch post says SFX v3 is available in Flows, the Sound Effects tab, Studio and the Music Editor, plus the ElevenLabs mobile app. Open the Sound Effects tab, type a description and generate.
In the API, set model_id to eleven_text_to_sound_v3 on the Sound Effects endpoint. If your account doesn't have access, the docs say you'll get an "Unknown model" error. Otherwise, this is what you can set:
| Setting | What it does |
|---|---|
| Duration | 0.1 to 30 seconds, or leave it blank and the model picks a length from the prompt |
| Prompt influence | High sticks closer to your words, Low gives the model more room |
| Loop | Set loop to true for a sound that repeats without a seam |
| Output | MP3, or WAV at 48 kHz for sounds that don't loop |
Example prompts to start with
These come from ElevenLabs' own docs:
| Prompt | What it tests |
|---|---|
| Rain against a window in a quiet apartment | A single ambient bed |
| Footsteps on gravel approaching slowly | Movement and distance |
| Football crowd cheering, hearing a ball drop, crowd booing | A sequence of sounds, the thing v3 was built for |
| Gospel choir singing hello | Vocals, one of the listed v3 improvements |
The docs have two writing tips. Word choice changes the result: "a couple arguing in French" came out better than "a couple fighting in French." And exact counts don't always work, so a prompt like "triangle hit 4 times" may not give you four hits. If you need a precise count, generate one hit and repeat it in an editor.
What it costs
ElevenLabs hasn't published a separate price for v3. The current sound effects docs charge 40 credits per second when you set a duration, and the pricing page FAQ puts a generation at roughly 200 credits. The free plan includes 10,000 credits a month, so about 50 generations at that rate, if your account has v3 at all. A commercial license starts at the Starter plan, so check your plan before you put a generated sound in a paid ad.
My read
Most small businesses won't need a sound effects model on its own. Where I'd use it is short video. A 15-second reel for a restaurant or a pool company usually goes out silent or with a stock track everyone has heard, because nobody had an hour to dig through sound libraries. If one prompt like "sizzling pan, plates set down, quiet dinner chatter" gets you a usable bed, that's the step people skip and now don't have to. I'd still listen to every clip before it goes out. The docs admit counts and exact wording trip it up.
ElevenLabs shipped Eleven v4 for voices last month, and video models like FLUX 3 Video now generate sound alongside the picture. Put those together and a small shop can make the voiceover and the sound for a promo from text prompts.
For developers, the limited-access note matters most. Don't build a workflow around eleven_text_to_sound_v3 until it shows up on your account, and keep v2 as the fallback.
If you're wondering which content or marketing tasks in your week AI could take over, New Face Design's free process audit is a good place to start. We'll look at where your hours go and point out the few jobs worth automating first.