The short version#
On September 28, 2026, ElevenLabs released Eleven v4, its new text-to-speech model, together with the low-latency Eleven v4 Turbo (ElevenLabs). Both are available now in the app, in ElevenAgents and via the API, and both speak German.
Two days after launch I tested both against the previous Eleven v3 with a German script full of traps: a date, a euro amount, a time, umlauts and long compound words. The result in one sentence: all three models read the script without a single error, but v4 generates it twice as fast as v3, and v4 Turbo is another three times faster.
If you want to try it: until October 12, 2026, a large share of v4 usage on Creator plans and above does not count against your balance (details below).
What's new in Eleven v4#
According to ElevenLabs, v4 is built on an entirely new architecture. The key points from the announcement:
- More expression: the model is meant to infer tone, pacing, emotion and context from the text. ElevenLabs cites the #1 spot on Artificial Analysis' voice leaderboard and roughly 75% preference in its own blind tests against competing models. Both are vendor claims.
- Direction by tag and plain language: you can describe the delivery in natural language or use tags like
[laughs]or[light rain]. v4 is said to follow these instructions more accurately. Custom pronunciations via IPA phonemes are also more reliable, per ElevenLabs. - Dialogue: multiple speakers respond to each other instead of isolated lines being stitched together.
- 90+ languages: a voice recorded in one language speaks others with a native accent while keeping its identity, and no longer drifts back to its source accent.
- Voice cloning: Instant Voice Clones are said to reach high fidelity from just 10 seconds of audio. Professional Voice Clones are supported in v4 as well.
- Long-form: chaining generations for audiobooks and long projects is significantly more reliable.
Eleven v4 Turbo is built for voice agents. ElevenLabs reports a median of about 150 ms to first audible speech, measured over WebSocket streaming with network latency removed, and about 100 ms of pure inference time.
My test: a German script, three models#
Setup: on September 30, 2026 I generated the same text with Eleven v3, Eleven v4 and Eleven v4 Turbo, using the same German voice from the ElevenLabs library ("Lars") with default settings, through the REST API with streaming from Germany. Each model ran three times; the table shows the median. To check pronunciation, I had Whisper ("medium" model) transcribe every recording.
The script (German, as tested):
Am 28. September 2026 hat ElevenLabs das Modell Eleven v4 veröffentlicht. Die Kfz-Haftpflichtversicherung übernimmt 1.250,50 Euro für den Schaden. Frau Müller aus Überlingen prüft das Angebot bis zum 12. Oktober um 14:30 Uhr. Übrigens gilt die Datenschutz-Grundverordnung auch für synthetische Stimmen.
| Model | Time to first audio | Full sentence generated | Recording length | Whisper transcript |
|---|---|---|---|---|
| Eleven v3 | 0.98 s | 15.8 s | 21.9 s | error-free |
| Eleven v4 | 1.55 s | 7.9 s | 18.9 s | error-free |
| Eleven v4 Turbo | 0.63 s | 2.5 s | 19.1 s | error-free |
What I take from it:
- Pronunciation: the date, amount, time, umlauts and the long compound "Kfz-Haftpflichtversicherung" came out so cleanly that Whisper returned every word correctly for all three models. For clean German, v3 is already enough; v4 is not a measurable quality jump here.
- Speed: v4 needs half as long as v3 for the whole sentence, v4 Turbo only a sixth.
- First audio: v4 starts slightly later than v3. For voice agents, v4 Turbo is the model to use; it delivered first audio in just over 0.6 seconds in my test.
- Speaking pace: v4 and v4 Turbo read the same text about three seconds faster than v3.
My latency numbers are not comparable to ElevenLabs' 150 ms: I measured over REST from Germany including the network, not over WebSocket. And one sentence with one voice is a spot check, not a benchmark. Whether v4 sounds more expressive is best judged by ear:
What ElevenLabs costs#
Prices from elevenlabs.io/pricing (as of September 30, 2026, monthly billing, excluding taxes):
| Plan | Price per month | Credits per month | Notable |
|---|---|---|---|
| Free | $0 | 10,000 | No commercial license (starts at Starter) |
| Starter | $6 | 30,000 | Commercial license, Instant Voice Cloning |
| Creator | $22 (first month $11) | 121,000 | Professional Voice Cloning |
| Pro | $99 | 600,000 | Higher-quality audio via API |
| Scale | $299 | 1.8M | 3 workspace seats |
| Business | $990 | 6M | 10 seats, 10 Professional Voice Clones |
Offer until October 12: on Creator plans and above, v4 usage of up to twice your monthly text-to-speech credits does not count against your balance, in the web and mobile apps only, not via the API. ElevenLabs markets this as "3x credits". If you have an audiobook or a voiceover series planned, these two weeks are the time to do it.
What else happened in August and September#
- Dubbing v2 (August 6): the new API dubbing preserves the tone, emotion and delivery of the original across 90+ languages and distinguishes regional variants such as Castilian and Latin American Spanish (ElevenLabs).
- Universal Music Group (September 10): ElevenLabs' first major-label agreement, covering licensing and a planned platform where fans can create remixes and mashups with music from participating artists (ElevenLabs).
- Music v2.5 (September 11): the new default model in ElevenMusic. The Free plan gets five lossless downloads per day with commercial use allowed if you credit ElevenMusic; Pro gets 400 lossless downloads per month (ElevenLabs).
- MCP for Claude, ChatGPT and Cursor (September 14): through the ElevenLabs connector, AI assistants generate speech, transcripts, dubs, music, sound effects, images and video right in the chat, with no server to run and no API keys to manage (ElevenLabs).
- Studio 4.0 (September 21): the video editor in ElevenCreative generates video, images, voice, music and sound effects straight onto the timeline, with 10,000+ voices in 32+ languages and a Studio Agent that drafts a first cut (ElevenLabs).
- ElevenLabs for Students (September 25): university students aged 18+ in the EU, US, Canada, Australia and the UK get three months of free access to ElevenCreative, ElevenAgents and the API, plus a free year of ElevenReader Ultra (ElevenLabs).
For freelancers and small teams in the DACH region#
ElevenLabs bills in US dollars, excluding taxes; what that means for VAT and reverse charge is covered in our post on AI tools as business expenses. And if you produce ads, explainers or podcasts with v4 for the EU market, the transparency duties of Article 50 of the AI Act apply, especially when a voice resembles a real person. We summarized the practice in Labeling AI voices: Article 50 in practice. This is not legal advice.
Which model for whom?#
- Audiobooks, voiceovers, explainers: Eleven v4. Faster than v3, with the new direction controls.
- Voice agents, phone bots, real-time apps: Eleven v4 Turbo. By far the fastest in my test.
- Existing v3 projects: no need to switch if pronunciation is right. For new recordings, v4 is worth it for the time savings alone.
- Beginners: the Free plan is enough to try it; for commercial use and your own voice clones you need at least Starter at $6. More on all features in the ElevenLabs guide, alternatives in the comparison with Murf.
FAQ#
How good is ElevenLabs in German?#
In my test, Eleven v3, v4 and v4 Turbo all read a German sentence with a date, an amount, a time and umlauts without errors. According to ElevenLabs, v4 keeps a native accent across 90+ languages.
How much does ElevenLabs cost?#
From $0 (Free, 10,000 credits) through $6 (Starter), $22 (Creator) and $99 (Pro) up to $990 (Business) per month, excluding taxes. The first Creator month currently costs $11.
Is ElevenLabs free?#
Yes, the Free plan includes 10,000 credits per month. For commercial use and Instant Voice Cloning you need at least the Starter plan.
What is the difference between Eleven v4 and v4 Turbo?#
v4 is tuned for expression and recording quality, v4 Turbo for low latency in voice agents. In my test, v4 Turbo was about three times faster than v4.
Sources#
- ElevenLabs: Introducing Eleven v4 (Sep 28, 2026)
- ElevenLabs: Pricing
- ElevenLabs: Dubbing v2 via ElevenAPI (Aug 6, 2026)
- ElevenLabs: Universal Music Group (Sep 10, 2026)
- ElevenLabs: Music v2.5 (Sep 11, 2026)
- ElevenLabs: Voice, music, image and video in the MCP (Sep 14, 2026)
- ElevenLabs: Studio 4.0 (Sep 21, 2026)
- ElevenLabs for Students (Sep 25, 2026)
- Own measurement on Sep 30, 2026: three runs per model, REST streaming, Whisper medium
