ElevenLabs drops Eleven v4
ElevenLabs says Eleven v4 can clone a voice from ten seconds of audio, but its own cloning guide still asks for at least a minute, and on cloned voices v4 ranks second on Artificial Analysis.
ElevenLabs released Eleven v4 on Monday, and the line everyone repeated was that an Instant clone now needs just ten seconds of audio. The company's own cloning guide, read on Tuesday, still tells you to record at least a minute and recommends one to two.
The acting gain looks real. On Artificial Analysis' voice arena, which scores each model on its own stock voices, v4 went from 17th (as v3) to first. When every model reads the same eight cloned voices, the test that matters if you clone yourself or an actor, v4 is second, behind Alibaba's Qwen-Audio-3.1.
The creative change is direction. Tags such as [whispers], [laughs] or [voice breaking] sit inside the script and steer each line, so the performance is only as good as the notes a writer puts on the page.
In practice, clone from a clean minute of your own voice, as the guide says, and spend the time you save writing direction into the scene. The launch price is $22 per million characters until 12 October, then $80.
Source: ElevenLabs blog and docs, Artificial Analysis, 28 and 29 Sep 2026.
