Release Summary
ElevenLabs launched Eleven v4 and Eleven v4 Turbo on September 28, 2026. They are two variants of one new text-to-speech model family, rather than consecutive upgrades. Eleven v4 prioritizes voice quality and expressive long-form work; Turbo applies the v4 technology to conversations where response time matters. This page covers the shared release and the choice between its variants.
The practical question is whether the audio is a finished asset or part of a live interaction. For an audiobook, game character, ad, or localized narration, start with v4. For a voice agent or interactive character, test Turbo against the application's full response-time budget. Both are available across ElevenCreative, ElevenAgents, and ElevenAPI, but those products have different billing units.
What Changed From Eleven v3
More faithful voices and expressive dialogue
ElevenLabs says v4 improves output quality, emotional delivery, audio-tag control, speaker consistency, and the accuracy of cloned voices over Eleven v3. It adds full Professional Voice Clone support and improves Instant Voice Clones. The company also reports more reliable stitching of separate generations for longer work, which matters when producing an audiobook or editing a long script in sections. These are vendor claims, not a guarantee that every existing voice will sound preferable after switching.
The model family supports more than 90 languages, compared with more than 70 for v3. The model reference lists a 10,000-character limit for a v4 request versus 5,000 for v3. Do not assume the same per-request limit for Turbo without checking its current endpoint documentation.
A real-time variant built for agents
Eleven v4 Turbo is tuned with ElevenAgents for live speech. ElevenLabs lists approximately 100 ms median model inference latency, excluding application and network time. Its launch announcement separately reports roughly 150 ms median time to first audible speech in its specified WebSocket test. These measure different stages; neither is an end-to-end caller response guarantee. A production voice agent still needs to account for transcription, reasoning, network, playback, and interruptions.
Eleven v4 or v4 Turbo?
| Decision | Eleven v4 | Eleven v4 Turbo |
|---|---|---|
| API model ID | eleven_v4 |
eleven_v4_turbo |
| Primary job | Finished narration, character performance, localization, and other quality-first audio | Real-time voice agents and interactive experiences |
| Main trade-off | Highest fidelity and expressive control in the family | Similar expressive direction with latency optimization |
| Languages | More than 90 | More than 90 |
| Regular API list price | $0.08 per 1,000 characters | $0.04 per 1,000 characters |
| September 2026 introductory API price | $0.022 per 1,000 characters through October 12 | $0.011 per 1,000 characters through October 12 |
The prices above are the ElevenAPI published character rates, checked September 29, 2026. They are not the same as ElevenCreative subscription credits or ElevenAgents call-minute charges. The introductory discount is time-limited; budget a continuing workload against the regular rate unless the current pricing page says otherwise.
For a one-million-character API job, the published list rates imply approximately $80 for v4 or $40 for Turbo before other charges and any account-specific allowances. This is a simple character-rate illustration, not a quote for an entire creative project or agent deployment.
Migration and Compatibility Checks
Switching the model ID is only one part of a v3 migration. The Eleven v4 guide identifies several changes worth testing:
- Compare your actual cloned voices. v4 follows the source recording's timbre, cadence, accent, and imperfections more closely. A voice may sound materially different from its v3 output, even when v4 is more faithful to the recording. Clean source audio matters.
- Test cross-language accents. When the output language differs from the reference voice's language, v4 aims for a natural target-language accent instead of retaining the original accent. This can improve localization but change a deliberately accented character voice.
- Review controls and prompts. v4 supports Stability, Similarity, and audio tags. Style and Speed sliders are unavailable, and SSML is unsupported. Audio-tag adherence is still being improved.
- Check Professional Voice Clones. Existing PVC workflows may need a v4 fine-tuning step in My Voices. Compare results with your own source material before moving a long-running production.
- Test the correct interface. Official model documentation lists both variants for Text to Dialogue WebSocket use. Verify the chosen model ID and endpoint in the current API docs rather than assuming an older integration accepts every v4 control.
ElevenLabs says v4 will continue to evolve after launch. Keep a small, repeatable set of voice samples and re-test critical accents, names, tag prompts, and latency as the model changes.
Who Should Switch, and Who Should Wait?
- Choose v4 if the output is edited or published audio and voice identity, multilingual performance, or nuanced character delivery is the main constraint.
- Choose v4 Turbo if the audio is generated inside a live conversation and you can measure the whole agent's response time, not just the model's inference step.
- Keep a v3 comparison if listeners prefer an existing clone's familiar delivery or if cross-language accent retention is essential. If your workflow depends on SSML or Style/Speed controls, verify which model and interface supports those required controls before switching.
- Consider Flash v2.5 separately when raw model latency matters more than the v4 family's expressive delivery; ElevenLabs still lists Flash as a distinct option in its model selection guide.
