# Four Voice Releases in Three Weeks: Eleven v4, Grok Voice 3.0, Fish Audio Drama 3 and Voxtral 3 Compared

- Published: 21 October 2026
- Updated: 21 October 2026
- Author: RenderBob team
- Category: Guides
- Canonical: https://renderbob.ai/blog/voice-model-releases-compared-eleven-grok-fish-voxtral

Four voice models have launched since 28 September: Eleven v4, Grok Voice 3.0, Fish Audio Drama 3, and Voxtral 3. Their headline numbers measure different steps.

Four voice models have launched since 28 September, each aimed at a different job. Their headline numbers are measured differently, so a lower figure on one row does not mean a cheaper or faster project.

## What each is for

- Eleven v4 and v4 Turbo, ElevenLabs, 28 September. Produced audio, and a real-time model for agents. Hosted. ElevenLabs reports more than 90 languages, and about 150 ms to first speech for Turbo. Our 13 October piece did not restate a starting plan price.
- Grok Voice 3.0, xAI. Speech-to-speech agents, via API, at $0.08 per minute, with time to first audio reported at 0.81 seconds. SiliconANGLE dates the release to 14 October, which is the date in our 16 October piece. The launch pages we used did not print a day, so treat that date as reported rather than as xAI's own dateline.
- Fish Audio Drama 3, 13 October. Directable text-to-speech. Open weights, hosted apps, and an API, with up to 16 speakers. A ComfyUI Wiki write-up lists 83 languages. Fish's own post, in the piece we published on 19 October, gave no language count.
- Voxtral 3, Mistral, 20 October. Transcription, text-to-speech, and an agent SDK. API, with self-hosting described as something you arrange with Mistral. Transcription from $0.0035 to $0.008 per minute, and text-to-speech at $13 per million characters.

## The numbers measure different steps

Time to first speech for text-to-speech, time to first audio for speech-to-speech, and transcription latency are different clocks. Prices are per minute, per character, or per plan. Line them up only after you convert to the unit your scene actually consumes.

## What matters for production

- Direction control. Eleven v4 uses inline audio tags, and Fish Audio uses plain-language directions in brackets. Test the same line on both.
- Hosting. Fish Audio ships weights you can run yourself. Mistral's page offers self-hosting by arrangement and does not state a licence. ElevenLabs and xAI, in the pages we have covered, are hosted services.
- Consent. ElevenLabs says cloned voices require verified owner consent. The Grok and Voxtral pages we reviewed describe no consent step. Fish's announcement did not discuss cloning. The Hugging Face licence we covered on 19 October bans cloning a real person's voice without consent. For real performers, get written permission whatever the tool says.
- Languages. The counts differ, and a count is not a listening test.
- Transcripts. Of these four, only Voxtral lists transcription as a core product.

## A shared test

Record one 30-second scene with two characters, including a name, a number, and a mid-line change of mood. Generate it on each model you can access and score pronunciation of the name and the number, the tone change, and consistency across three runs. When that voice goes into a picture cut, RenderBob lets you choose local or cloud for each clip and keep the references with the studio.

## What to watch

Independent listening tests, consent policies for Grok and Voxtral, and a licence statement for Voxtral 3.

## RenderBob 3 and this story

RenderBob 3 applies the same habit to picture that this comparison applies to voices. Try a model on one shot, keep the take that works, and the choice stays on the shot record.