Eleven v4

Not yet independently verified. Every detail here comes from ElevenLabs’ own announcement and model page; no independent report or measurement of Eleven v4 was found. The latency figures and the listener-preference result are ElevenLabs’ own, and its pages give no per-character API price. We will update this when it can be confirmed, and remove this note.

Eleven v4 is the text-to-speech model ElevenLabs announced on 28 September 2026, in a post by its co-founders Mati Staniszewski and Piotr Dabkowski. The company says it is built on a new architecture that reads a script the way a voice actor would, taking account of who is speaking and what has just happened, and generates speech in more than 90 languages. Delivery can be directed in plain language or with inline tags for emotions, accents and sound effects, which ElevenLabs says v4 follows more reliably than its v3 model; pronunciation through International Phonetic Alphabet phonemes is also said to be more reliable. Context stitching keeps pacing steady across long scripts such as audiobooks, regenerating a line keeps the same voice, and Professional Voice Clones, which v3 did not support, work again. It is available in ElevenLabs’ creative tools, its ElevenAgents platform and its API; it is a hosted model with no published weights, and the pages we read give no per-character API price. ElevenLabs cites a third-party ranking and its own blind listening tests in its favour; we have not checked either, so no score is stored. A low-latency variant, Eleven v4 Turbo, was released at the same time; the board also lists the earlier ElevenLabs Multilingual v2.