ElevenLabs: Eleven v4 gives AI voices emotion, and Turbo starts speaking in about 150 ms

What's new

On September 28, 2026, ElevenLabs launched two new voice models: Eleven v4 and Eleven v4 Turbo. Both convert text to speech. v4 is built on an entirely new architecture that, according to ElevenLabs, reads a script much like a voice actor would, putting more emotion into the voice. The model supports more than 90 languages, multiple speakers in the same recording and built-in sound effects.

v4 Turbo is the company's fastest real-time model and offers the same expressiveness. It is available through the API (the technical gateway developers use to connect services) and in the ElevenAgents platform.

What's smart

You can steer the delivery with small tags in the script, such as [laughs], [whispers] or [door slams]. ElevenLabs claims that v4 follows strings of such tags more reliably than its predecessor, v3.

Professional voice clones, which could not be used in v3, are back. They can be used with the model's full emotional range and sound the same on v4 and v4 Turbo, so a company's brand voice can stay consistent across both. According to ElevenLabs, the context stitching feature is meant to keep pacing and delivery stable throughout long scripts, such as an entire audiobook.

Turbo offers bidirectional streaming: audio can start coming back before a sentence is finished. It is designed for agents, where a large language model (LLM) writes the response while the voice is already reading it aloud.

Cheaper or better?

On speed, ElevenLabs states a median inference latency of about 100 ms for Turbo and a median time to first speech of about 150 ms. In a comparison reported by the outlet DataNorth AI, Cartesia Sonic 3.6 came in at 262 ms and OpenAI GPT-4o mini TTS at 814 ms. The measurement does not come from an independent test, and the methodology is unknown, so the figures should be taken with a grain of salt.

On price, according to DataNorth, ElevenLabs' subscriptions start at $6 per month, and there is a free tier with 10,000 credits per month. The models are available in the web app, but the sources do not confirm that v4 can be tried for free. DataNorth found no price difference between v4 and v4 Turbo. API prices per million tokens and availability in Denmark and the EU are not stated in the sources.

What it's good for

v4 is positioned for prerecorded audio such as audiobooks and voiceovers. The changelog also describes it as a model for high-quality dialogue via the Text to Dialogue API. Up to 10,000 characters can be generated at a time, and the existing roughly 17,500 voices can be used.

v4 Turbo is for real-time use: phone agents and live assistants, where long pauses ruin the conversation. In a quote on ElevenLabs' site, Salesforce says Turbo delivers faster and more natural responses for Agentforce Voice.

For an ordinary business, this means a customer service voice can sound more lifelike and respond faster. For creators, an audiobook or a video can get a narrator with feeling, without a recording studio.