Eleven v3

Eleven v3 is ElevenLabs' third-generation text-to-speech model, launched in alpha June 2025 and generally available from February 2026. Its signature is an inline markup language of bracketed audio tags — [laughs], [whispers], [excited], [sighs] — that direct emotion, delivery, and non-verbal reactions inside the script itself, plus native multi-speaker dialogue with interruptions and emotional transitions via the Text to Dialogue API. It supports 70+ languages and ranks near the top of the Artificial Analysis Speech Arena.

The honest picture: v3 trades latency for fidelity — ElevenLabs itself steered real-time use cases to the Flash family (75ms vs v3's much higher latency); a dedicated v3 Conversational variant later closed that gap for voice agents. Alpha pricing was discounted, but steady-state costs run higher per character than older models. Character limits (5,000/request) make long-form narration a chunked workflow.