This update brings new voices, music and image models and five chat models to ToRun, and lets the voices that act out audio tags hear them. Here is what is new and how each one is used.
Audio tags reach the voice
Some voice models act out short English cues written in square brackets, such as [laughs], [whispers] or [sighs]. Write “That was close [laughs]” and the voice laughs instead of saying the word.
ToRun used to remove these cues from your text in most places before the voice model saw them. They now reach the voice models that perform them, Eleven v3 and Eleven v4 Turbo, in Text-to-Speech, when the chat reads an answer aloud and when you ask the chat for speech. Other voice models are given the text without the cues, so they never read a cue aloud, and the same is true when a request moves to another voice model. In Text-to-Speech, a note under the text box says when the chosen voice model acts out cues, with examples. Write the cues in English, whatever language the rest of the text is in.
New voices
- Eleven v4 is ElevenLabs' most expressive voice model. It costs the same per character as Eleven v3.
- Eleven v4 Turbo is its faster sibling. It acts out audio tags and costs the same per character as Flash v2.5.
New music models
- Eleven Music v2.5 costs the same per minute as Eleven Music v2.
- Stable Audio 3 makes songs of up to five minutes, longer than Stable Audio 2.5, at the same price per song.
ElevenLabs retires Eleven Music v1 on 31 January 2027. A song set to Eleven Music v1 is now made with Eleven Music v2.5, at the same price per minute. Music and audio generation explains how a song is made and charged.
New image models
- Recraft V4.1 costs a little less per image than Recraft V4.
- Recraft V4.1 Flash costs much less. In our tests it did not write every word inside an image exactly, so check the text in its pictures, or choose another model for pictures with words in them.
Five new chat models
- Grok 4.7 from xAI, at the same price as Grok 4.6.
- GPT-6.1 Terra from OpenAI, the middle model of its GPT-6 line.
- Claude Haiku 5.5 from Anthropic, the lowest-priced Claude model in ToRun.
- Gemini 4 Argon from Google. Google charges a launch price until 30 November and twice as much from 1 December, and ToRun follows on that day.
- Mistral Large 4 from Mistral.
Grok 4.7, GPT-6.1 Terra and Claude Haiku 5.5 also have a long-context version in the model list. Choose it for very long conversations. It costs more, because the provider charges a higher rate for long prompts.
Mistral names Mistral Large 4 as the successor of Mistral Large 3 and Mistral Medium 3.5. ToRun's requests for these two models were being refused, so a chat set to one of them now runs on Mistral Large 4 and is charged its price. A chat set to Codestral 2 now runs on Codestral 2508.
How Auto treats a new model
When you leave the choice of model to Auto, it chooses among the models that have been checked for the job. A new model joins Auto only after we have listened to its voice, looked at its pictures or tried it in conversations. In this update, Eleven v4, Eleven v4 Turbo, Eleven Music v2.5 and Recraft V4.1 have joined Auto, and the models Auto prefers stay the same. Auto does not choose the five chat models, Recraft V4.1 Flash or Stable Audio 3 for now. You can choose any of them yourself. Model routing and Auto explains how Auto chooses.
Also in this update
- When xAI refuses a video for its content, ToRun now tells you at once, instead of reporting after ten minutes that the video took too long. The charge is returned, as before.
- When a photo edit makes several pictures, every one of them is saved to your Library.
- In a meeting call, live captions keep following the newest line.