FULL STORY

ElevenLabs v4: Launch, Hands-On Praise, and Platform Rollout

ElevenLabs launched Eleven v4 and v4 Turbo on September 28, topping TTS leaderboards. Hands-on reviews praised its realism and accent removal, the model landed on Runway, and free access runs through October 12.

2026-09-28 ~ 2026-10-05 · 5 episodes · 33 posts

Episode 1 · ElevenLabs Launches Eleven v4, Tops TTS Leaderboards at a Premium Price (2026-09-28, 22 posts)

On September 28, ElevenLabs launched its next-generation text-to-speech models, Eleven v4 and v4 Turbo, describing them as its fastest and most emotionally expressive to date. Independent benchmarks from Artificial Analysis show Eleven v4 topping both the Provider Voice Arena and the Pronunciation Robustness benchmark, cementing its lead in TTS—though at $80 per million characters, roughly five times the price of Google's competing offering, the performance-versus-cost tradeoff is a key consideration for users.

Confirmed

  • ElevenLabs officially released Eleven v4 and v4 Turbo; v4 uses a new architecture to generate speech with intended tone, pacing, emotion and character, aimed at developers building creative experiences, while v4 Turbo focuses on low latency (@testingcatalog, @NevFlynn)
  • The team shipped an interactive demo showing word-level emotional control via [audio tags] such as [excited], [playful] and [amazed] in prompts (@NevFlynn)
  • Scripts can embed performance directions directly: tags like [whispers], [laughs], [long pause], [door slams] control emotion, pauses, laughter, whispering, actions and ambient sound effects, and natural language can describe how a line should be delivered—no repeated tuning or batch-and-pick needed; voice cloning works with roughly 10 seconds of sample audio (@xiaohu)
  • Developer korymath noted v4 lets users direct both character performance and surrounding sound environment, and supports voice effects such as a "cheap microphone" timbre (@korymath)
  • Developer khademinori added that v4 introduces IPA phonetic support for precise pronunciation control, with more natural multi-speaker dialogue (@khademinori)
  • ElevenLabs member lukeharries demoed upgraded voice cloning, calling it the company's fastest, most expressive model, now live in ElevenCreative, ElevenAgents and ElevenAPI (@lukeharries)
  • User thealexbanks reported being unable to distinguish their cloned voice from real recordings (@thealexbanks)
  • First place in the Provider Voice Arena with Elo 1,319 across 1,674 appearances, ahead of Cartesia Sonic 3.6 (1,276) and Gemini 3.8 Flash TTS (1,267), with Alibaba's Qwen close behind (@ArtificialAnlys)
  • Topped the Pronunciation Robustness benchmark covering four categories including abbreviation expansion, scoring above 93% in three of them, judged line-by-line by human raters against agreed reference pronunciations; supported languages expanded to 90+ (@ArtificialAnlys)
  • Generation speed of 73.4 characters/second versus 42.5 for v3; priced at $80 per million characters, roughly five times Google's competing offering (@ArtificialAnlys)
  • Artificial Analysis shared two sample clips—a customer-service script and a black-hole explainer—showing real-world performance (@ArtificialAnlys)
  • Third-party user mati shared the evaluation on X, calling Eleven v4 "clearly number one" in comparisons (via @annbordetsky)

Why it matters

  • Topping both the arena and pronunciation benchmarks while nearly doubling generation speed, and shipping across ElevenCreative, ElevenAgents and ElevenAPI, Eleven v4 consolidates ElevenLabs' lead in TTS
  • The new architecture's emotion tags, script-embedded performance directions, voice effects, IPA pronunciation control and fast cloning push TTS from mere narration toward directable creative production
  • At $80 per million characters—about five times Google's price—cost-versus-performance is a key selection consideration, and Cartesia, Gemini and Alibaba's Qwen remain close competitors

2 more related posts →

Episode 2 · ElevenLabs v4 Early Tests Praise Unprecedented Voice Realism, Accent Removal (2026-09-29, 3 posts)

Early hands-on tests of ElevenLabs v4 are overwhelmingly positive, with users calling the voice realism unprecedented. One tester found v4 nearly removed her accent when cloning her own voice.

Episode 3 · ElevenLabs v4 Voice Model Now Available on Runway (2026-09-29, 2 posts)

Runway announced that the ElevenLabs v4 voice model is now available on its platform, offering notably improved expressiveness and natural intonation for narration and character dialogue, and can be used alongside Runway's video tools.

Episode 4 · ElevenLabs launches Eleven v4 voice model, free until October 12 (2026-10-01, 4 posts)

ElevenLabs has launched Eleven v4, a new architecture that performs text with real emotion rather than just reading it, expanding language support from 32 to 90+ in ElevenReader. It is free to use until October 12, with bonus quotas for paid plans.

Episode 5 · Eleven v4 Tested: Zero-Edit Outputs and Natural Speech Slowing (2026-10-04, 2 posts)

Early tests of ElevenLabs' free Eleven v4 voice model show it is the first to produce usable output with zero edits, and it now responds naturally to the [slow] tag, slowing speech without sounding dull.