ElevenLabs launched SFX v3 on October 29, 2026 — a new text-to-sound-effects model aimed at scenes that need more than one sound in a single clip. If you make short video, game audio, or social Foley, this is the update that matters: one prompt can ask for a crowd, a ball drop, and a boo in sequence, instead of stitching three generations by hand.

Coffee summary

  • FACT: SFX v3 is available in ElevenCreative (Sound Effects, Flows, Studio, Music Editor) and via ElevenAPI with model_id = eleven_text_to_sound_v3.
  • FACT: Docs say v3 follows multi-part prompts more closely, handles music/vocals better, and is less likely to clip on high-frequency sounds than eleven_text_to_sound_v2.
  • FACT: Max duration is 30 seconds per generation; optional duration billing is 40 credits/second when you set length explicitly.
  • FACT: SFX v3 is in limited access and may not be available in every region — accounts without access get an Unknown model error.
  • CLAIM (ElevenLabs): v3 produces "cleaner, more natural audio than SFX v2 in benchmark testing." Benchmarks are not published in the launch post, so treat that as a vendor claim until you A/B your own prompts.
  • OPINION: For creators, the win is prompt structure, not magic quality — write sequences, not vague categories.

What actually shipped

The blog post frames the problem clearly: a football clip might need cheering, a ball drop, and booing seconds later. SFX v3 is built to keep those parts in one generation. The docs add the practical controls you need in production:

Control What it does
eleven_text_to_sound_v3 New model; v2 remains the default (eleven_text_to_sound_v2)
Duration 0.1–30s; auto if omitted
Looping Seamless loops for rain, ambience, beds
Prompt influence High = literal; Low = more variation

Output is MP3 for all effects; WAV at 48 kHz for non-looping effects.

Prompt pack (copy/paste)

Use action + setting + atmosphere. Avoid "hit 4 times" style counting — docs say exact repetition counts are unreliable.

Video / social
1. Busy cafe interior, espresso machine hiss, ceramic cup set on saucer, quiet chatter under soft jazz bed
2. Phone notification ping, then fabric rustle as someone picks up the phone, distant traffic through a window
3. Door slam in a hallway, footsteps jogging away on tile, elevator chime arrives
4. Keyboard clacking close-mic, mouse click, chair roll on hardwood, HVAC hum under it
5. Rain against a window in a quiet apartment at night, occasional car pass outside

Game / interactive
6. Footsteps on gravel approaching slowly, metallic latch unlock, heavy wooden door creaks open
7. Sword drawn from scabbard, two blades clash, sparks fizz, then a short whoosh as the camera pans
8. Sci-fi UI confirm blip, soft servo whir, then a low drone under a loading chamber
9. Inventory pickup chime, leather bag zip, potion glass clink
10. Boss-room braam hit, debris rattle, distant crowd murmur behind a stone gate

Trailer / cinematic
11. Thunder rumbling in the distance, wind through dry grass, single crow call
12. Glass shattering on concrete, alarm starts after a half-second pause
13. 90s hip-hop drum loop, 90 BPM, dusty vinyl crackle under the kick
14. Vintage brass stabs in F minor, short trailer braam tail
15. Soft rain loop bed, then a door opens and a wet coat drops on the floor

Limitations you should plan for

  • Region / access: limited access; verify the model works on your account before promising it in a client workflow.
  • 30-second ceiling: longer beds need looping or the Music API for full tracks.
  • Unpublished benchmarks: quality claims are vendor-side; run your own A/B with v2 vs v3 on the same prompt set.
  • Exact counts: "triangle hit 4 times" style prompts are not reliable.

AIImpish Take

SFX v3 is useful when your scene is a sequence, not a single hit. Write prompts the way you write a shot list. If your account does not have access yet, keep a v2 fallback and the same prompt pack — the prompting skill transfers.

*Sources: ElevenLabs, Introducing SFX v3, Oct 29, 2026; Sound effects docs.*