ElevenLabs has launched Eleven v4, a new speech model designed for improved consistency and expressiveness. The architecture analyzes script tone, pacing, and context, allowing users to guide output through tags, plain sentences, or phonetic spelling. The model supports over 90 languages, an increase from the 70 supported by v3, and enables native accents in cloned voices.
The Turbo variant of Eleven v4 is optimized for real-time applications, such as customer service agents and gaming. It achieves a speech generation latency of 150 milliseconds, outperforming Cartesia Sonic 3.6 at 262 milliseconds and OpenAI GPT-4o mini at 814 milliseconds. The model is integrated with the ElevenAgents platform to balance speed and quality.
Eleven v4 handles up to 10,000 characters per request, equivalent to approximately ten minutes of audio. It improves pronunciation accuracy, scoring 91.7 percent on benchmarks compared to 85.6 percent for v3. The model is currently available via the API, ElevenAgents, and ElevenCreative, with data storage options available in the US, EU, India, and Singapore.
Standard API pricing for Eleven v4 is set at $80 per million characters, while the Turbo variant costs $40. Through October 12, promotional rates are $22 and $11, respectively. ElevenLabs also offers a marketplace where voice actors can license their clones to earn revenue, featuring celebrity voices such as Michael Caine.
