ElevenLabs’ SFX v3 can build a whole scene from one prompt

ElevenLabs has released SFX v3, its newest sound effects model, which the company says follows prompts that describe several sounds at once more reliably than SFX v2. It is available in ElevenCreative and through ElevenAPI.

One prompt, a crowd, a ball and boos

ElevenLabs opens its announcement with a football clip as the example. A scene like that needs a cheering crowd, a ball dropping and the same crowd booing seconds later, and until now getting all of it from a single prompt took trial and error, according to the company. SFX v3 is meant to include more of the sounds a prompt names.

The company lists three improvements over SFX v2, all based on what it calls benchmark testing. It does not publish scores, a test set or a method, so the comparison cannot be checked from the post.

• Multi-sound prompts: the model follows several instructions in one prompt, so more of the described sounds appear in the output.
• Music and vocals: prompts that include a musical phrase or a vocal performance are handled better, which ElevenLabs says allows effects that mix them with environmental and everyday sounds.
• Cleaner audio: overall quality is higher, with fewer unwanted artifacts on gongs, chimes and whistles. ElevenLabs gives the example of a whistle that keeps a steady tone instead of turning harsh or crackly.

Inside ElevenCreative, text-to-effect generation is offered in the Flows canvas, a dedicated effects tab, the Studio editor and the Music Editor. The company says SFX v3 also works in the ElevenLabs mobile app. Developers can reach it through ElevenAPI to build effect generation into their own products.

Access may be narrower than the launch post suggests

The blog post says SFX v3 is available now and gives no price, plan restriction, regional limit or maximum clip length. The ElevenLabs developer documentation for sound effects tells a more cautious story. As of this writing it lists eleven_text_to_sound_v2 as the default model and describes eleven_text_to_sound_v3 as being in limited access, possibly unavailable in some regions. According to the same page, requests from accounts without access return an “Unknown model” error.

That wording may simply lag the launch, and the two pages are not reconciled. Developers who want to use v3 should test it with their own account before planning a rollout. The blog also does not say whether v3 replaces v2 in the interface or sits alongside it, or whether the earlier model stays the API default.

Limits and costs that decide real use

The announcement is silent on numbers that matter for production work, so the existing documentation is the best guide, with the caveat that it was written for the sound effects feature in general and not specifically for v3. It says each generation can run up to 30 seconds, with a duration setting between 0.1 and 30 seconds. If a duration is specified, the page lists a cost of 40 credits per second. Longer effects can be made seamlessly loopable, and a prompt influence setting trades literal adherence for a looser, more varied interpretation.

At those figures, a full 30-second layered scene with a set duration would cost 1,200 credits per attempt. Since ElevenLabs itself describes a workflow of building from the main action outward and then refining the description, several generations per scene is the expected pattern. Whether v3 changes the credit rate is not stated, so the real cost of iterating is unknown for now.

The claim about reliability also needs context. Better adherence to multi-sound prompts reduces retries, which is where most of the cost sits, but only if the improvement holds outside the company’s own examples. The post includes audio demos, not measured results, and no independent evaluation has been published.

A sign of where ElevenLabs is spreading

The release fits a broader pattern at ElevenLabs, which has been extending from voice into adjacent audio. Only days ago the company described Eleven Realtime, a family of speech-to-speech models for voice agents. Its voice library is also an established product with an API. SFX v3 pushes the sound effects line toward the Music Editor and Studio workflows, where effects, music and vocals are edited side by side.

The focus on music and vocals inside an effects model is notable. It suggests ElevenLabs wants one prompt box to cover more of a video or game soundscape, so that creators spend less time moving between separate tools. That is an inference from the product description, not something the company states, and the post says nothing about pricing strategy or planned follow-up releases.

What the source does establish is narrow: a new model, three claimed quality gains over its predecessor, and the places where it can be used. What it leaves open is how large those gains are, who can call the model through the API today, and what a layered effect costs in credits.