One prompt, a whole scene
A football clip might need a crowd cheer, the sound of a ball dropping, and then a change in the crowd's reaction. Rather than generating each effect separately, SFX v3 is designed to interpret a prompt describing several sounds and produce them together in one generation. How accurately it follows the requested order is something users still need to test.
SFX v3 is ElevenLabs' attempt to skip the stitching. The pitch is a prompt that describes a scene — several events, in an order, in a place — and a single generation that lands them together. It sits inside the tools video and game editors already open: Flows, the Sound Effects tab, Studio and the Music Editor in ElevenCreative, the mobile app, and the API for anyone building it into a pipeline.
That scope is worth stating precisely, because another approach is emerging in the category. Mirelo's SFX 2.0, covered here in October, uses video as input to help align effects with the footage. ElevenLabs SFX v3 works from text prompts instead. It does not automatically inspect the edit, so users remain responsible for placing and synchronising the generated audio.
The guidance gives away the weak point
ElevenLabs' prompting advice favours action, setting, atmosphere and performance over broad category words. Its newer SFX v3 documentation encourages multi-part prompts, including sequences of events. But the broader sound-effects guidance still recommends generating complex sequences individually and combining them in an editor for the best results. The new model makes one-pass generation more practical; it does not eliminate the need to inspect and edit the result.
Then the admission: exact counts are still unreliable. Ask for a specific number of repeats — "triangle hit 4 times" is the example in the guidance coverage — and you may get a different number. For ambience or a loose sound-design layer, that may not matter. For a transition that must hit a precise number of times against an edit, it does. The documentation is effectively telling users not to rely on the model for exact repetition counts, particularly when timing and repeatability matter.
Cleaner, on their scorecard
The quality claims are narrower than most launch copy, and they are still the company's own.
ElevenLabs says v3 follows multi-sound instructions more reliably, handles music and vocals more convincingly, and is less likely to clip on high-frequency sounds such as gongs, chimes and whistles than v2. These are company-reported improvements based on internal benchmarks; the underlying benchmark tables have not been published. ElevenLabs itself says results vary by prompt and recommends comparing both versions using the prompts users actually need.
The benchmark tables are not published, and the secondary coverage reviewed for this article did not independently reproduce a v2-versus-v3 comparison. ElevenLabs' general sound-effects product page, checked on October 30, also made no mention of v3: no version name, comparison or scores. Its quality language describes the generator generally. That does not contradict the launch announcement, but it leaves readers without an independently inspectable product-page comparison. "Cleaner" is exactly the kind of claim that needs a real project test.
The honest test is unglamorous: take the house prompts you already use, run them through v2 and v3, and count the retakes. The model that makes you redo fewer takes wins, whatever a scorecard says.
API pricing is separate from the subscription plans. ElevenLabs' documentation specifies usage-based charging, including 40 credits per second when a duration is explicitly set. A secondary pricing reference puts the API list cost at approximately US$0.12 per minute, but that should be treated as an indicative figure rather than a universal rate. Actual costs depend on the account's API pricing arrangement.
The sound-effects product page lists monthly plans at US$0 for Free, US$6 for Starter, US$22 for Creator and US$99 for Pro, before taxes. Commercial-use rights are included on paid tiers, while Free is limited to personal use. Promotional prices and introductory discounts may differ.
And access is not universal on day one: the docs note that some regions or accounts get an "Unknown model" error, so a pipeline should not be switched over before the model actually answers on your account.
Where this leaves video sound
AI video models increasingly arrive with native audio of their own, which raises a fair question: who still needs a separate sound-effects model? The working answer is anyone whose clip needs a specific scene sounded in a specific order, after the generation is done. Native audio is whatever the video model decided. A text-to-SFX pass is the layer you direct.
V3 is designed to make that layer easier to generate in one pass rather than requiring users to build every sound separately. Whether it produces better results in real editing workflows remains unproven outside the company's own testing.
Sources
- [1] ElevenLabs, "Introducing SFX v3" launch video (primary — description read first-hand): SFX v3 available across ElevenCreative in Flows, the Sound Effects tab, Studio and the Music Editor, with the link to the sound-effects product pageRead source
- [2] ElevenLabs sound-effects product page (primary — read first-hand via live browser on Oct 30, 2026, including its pricing table and FAQs: 30-second max on all tiers, Free MP3 44.1 kHz / paid MP3 + WAV 48 kHz, seamless looping on all tiers, monthly plan prices, commercial use paid-only; the page nowhere mentions SFX v3, v2, benchmarks or a model_id — a dated observation, not evidence about availability)Read source
- [3] ElevenLabs sound-effects documentation and API reference, as documented in NVdium's Oct 30, 2026 close read of the company blog and docs (labelled honestly: this newsroom's text fetcher cannot read elevenlabs.io directly): prompting guidance, the exact-count limitation, usage-based API charging including 40 credits per second when a duration is set, and the "Unknown model" access caveatRead source
- [4] NVdium, "ElevenLabs SFX v3: Multi-Sound Prompts, Cleaner Audio", Oct 30, 2026 (read first-hand): close read of the company blog and docs — the quality claims, model_id eleven_text_to_sound_v3, the roughly $0.12-per-minute API list figure, the access caveat and the prompting guidanceRead source
- [5] The Neuron, "Everything That Happened in AI Today", Oct 30, 2026 (read first-hand): roundup listing the SFX v3 release and its ElevenCreative and ElevenAPI availabilityRead source