Eleven v4 Is Free to Try: ElevenLabs’ #1-Ranked AI Voice Model (Plus Open-Source Alternatives You Can Run Locally)

ElevenLabs shipped Eleven v4 and Eleven v4 Turbo on September 28, and within days Eleven v4 took the top spot on an independent text-to-speech leaderboard. It is the new model to beat for expressive voiceover: inline emotion tags, 90+ languages, and an instant voice clone that ElevenLabs says needs just 10 seconds of audio.

The catch for creators on a budget: Eleven v4 is closed and hosted, and the free plan is for personal use only. So this post does two things. It shows you how to try Eleven v4 free today in the browser, and it gives you the open-weight alternatives you can run on your own machine, with the licenses spelled out, for when you need commercial rights without a subscription.

Eleven v4 at a Glance

• What it is: ElevenLabs’ new flagship text-to-speech model (eleven_v4) plus a low-latency sibling (eleven_v4_turbo) for real-time agents.
• What’s free: the ElevenLabs free plan gives 10,000 credits a month, no card required, and lets you generate with Eleven v4 in the web app.
• What it costs: commercial use and instant voice cloning start on the Starter plan (normally $6/month, discounted at the time of writing). Via API, Artificial Analysis lists Eleven v4 at $80 per 1M characters.
• The ranking: #1 on the Artificial Analysis Provider Voice Arena at 1320 Elo, a clear 29-point lead over second place.
• The one catch: free-plan audio is not licensed for commercial work, and there are no open weights. For a free commercial path, use Qwen3-TTS (Apache 2.0) or Chatterbox (MIT).

What’s New in Eleven v4

According to ElevenLabs’ launch post, v4 is a new architecture rather than a v3 refresh. The changes that matter for creators:

• Inline direction tags. You can write cues like [laughs], [said angrily in French accent], [light rain] or [phone buzzing] straight into the script. The model handles both delivery and simple sound effects from the same text.
• 90+ languages, up from 70+ on Eleven v3, and a voice recorded in one language can speak the others.
• Longer generations. The models docs list a 10,000-character limit per request for v4 (about 10 minutes of audio), double v3’s 5,000.
• Better cloning. Instant Voice Clones from about 10 seconds of audio, plus Professional Voice Clone support for top-fidelity work.
• Improved IPA phoneme support, which helps with brand names, character names and other words the model usually gets wrong.

Is Eleven v4 Really the Best AI Voice? What the Leaderboard Says

ElevenLabs’ own numbers say about 75% of listeners preferred v4 in blind head-to-head tests, and that v4 Turbo reaches first speech in a median of ~150 ms versus 262–814 ms for the competitors it tested. Those are vendor claims. We have not been able to reproduce them, so treat them as marketing until someone else measures them.

The independent number is more useful. On the Artificial Analysis Provider Voice Arena, where listeners vote blind, Eleven v4 sits at 1320 ±18 Elo across 1,803 samples. Qwen-Audio-3.1-TTS-Plus is second at 1291 ±18. The confidence ranges barely touch, and the leaderboard ranks v4 alone at #1, so the lead looks real and not a statistical tie. The race for third is a different story: Sonic 3.6 (1273) and Gemini 3.8 Flash TTS (1272) are effectively tied.

That price column is the real story. Eleven v4 costs roughly four to five times as much per character through the API as Qwen-Audio-3.1 or Gemini 3.8 Flash TTS, which we covered last week. You’re paying a premium for about 30–50 Elo points.

Eleven v4 vs the Free and Open Alternatives

Try Eleven v4 Free (and the Open-Source Route)

Option 1: Eleven v4 in the browser (no GPU, personal use). Create a free account at elevenlabs.io/v4 (no credit card), then open the Text to Speech app. Choose Eleven v4 in the model dropdown, paste a script with a couple of tags (for example [whispers] Don't move. [pause] It's right behind you.), and generate. Watch the credit counter: 10,000 credits go fast on long scripts, so test with short lines first.

Option 2: Qwen3-TTS (free, Apache 2.0, commercial OK). The fastest way in is the free Qwen3-TTS demo Space, which runs on Hugging Face ZeroGPU and has three tabs: Voice Design (describe a voice in words), Voice Clone, and preset CustomVoice speakers.

To run it locally, the 1.7B CustomVoice model card lists about 2B parameters in BF16, which is roughly 4 GB of weights. By our estimate, an 8 GB NVIDIA GPU is a comfortable floor. Install and generate:

pip install -U qwen-tts

python -c "
import torch, soundfile as sf
from qwen_tts import Qwen3TTSModel
m = Qwen3TTSModel.from_pretrained('Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice', device_map='cuda:0', dtype=torch.bfloat16)
wavs, sr = m.generate_custom_voice(text='Welcome back to the channel.', language='English', speaker='Ryan')
sf.write('out.wav', wavs[0], sr)"

Source and docs are on GitHub (QwenLM/Qwen3-TTS). For prompt-based voice design, swap in the VoiceDesign checkpoint.

Option 3: Chatterbox (free, MIT, light hardware). Try the Chatterbox Multilingual V3 Space, or install it with pip install chatterbox-tts. At 0.5B parameters it’s the easiest of the three to run on a mid-range card, and it has an exaggeration slider for emotional intensity. See the model card for the full example. If you’d rather skip Python entirely, our guide to audio.cpp, the free engine that runs 62 audio models locally, covers a no-Python route for several open TTS models.

Honest Trade-offs

• Free plan is not a commercial license. If the voiceover goes in a client video or a monetized channel, you need a paid ElevenLabs plan or an Apache/MIT model.
• No voice cloning on free. Instant Voice Clone starts on Starter, so the “10 seconds of audio” headline doesn’t apply to a free account.
• API pricing is steep. At $80 per 1M characters, an audiobook-length project costs several times more than on Gemini or Qwen’s hosted models.
• The speed and preference numbers are self-reported. The 75% preference and ~150 ms first-speech figures come from ElevenLabs. Only the arena Elo is independently measured.
• The open models are a step behind. The top open-weights entry on the same arena, Breeze TTS 2, scores 1206, more than 100 Elo below v4, and it’s non-commercial. Expect flatter emotion from open models on dramatic reads.

Workflow Tips for Creators

• Draft free, finish paid. Block out the timing and tags for a script in Qwen3-TTS or Chatterbox, then spend Eleven v4 credits only on the final read.
• Tag sparingly. One or two cues per paragraph usually beats tagging every line. Too many tags can make the read sound performed.
• Write tricky names in IPA instead of regenerating five times. v4’s improved phoneme support pays off on brand and character names.
• Keep a license note per project. Record which model and plan made each clip, so a client question about rights takes ten seconds to answer. Our Tencent AuK breakdown shows how a hidden component license can trip you up.
• Use v4 Turbo only for live work. For recorded voiceover, the full v4 model’s quality matters more than shaving 100 ms.

Want to see an ElevenLabs cloning workflow end to end? Watch How to Clone Your Client’s Voice in Under 5 Minutes with ElevenLabs on the Creative AI Show YouTube channel. The same steps apply to v4 on a paid plan.

Eleven v4 FAQ

Is Eleven v4 free?

You can use it free. The ElevenLabs free plan includes 10,000 credits a month with no credit card required, and you can pick Eleven v4 in the web app. The free plan covers personal use only.

Can I use Eleven v4 commercially?

Only on a paid plan. Commercial rights start at the Starter tier. For free commercial use, run an Apache 2.0 or MIT model such as Qwen3-TTS or Chatterbox.

What are Eleven v4’s system requirements?

None locally. It’s hosted, so any modern browser works. The open alternatives need a GPU for good speed. Qwen3-TTS 1.7B has about 4 GB of weights (we’d plan on 8 GB of VRAM), and Chatterbox is smaller at 0.5B parameters.

What’s the difference between Eleven v4 and Eleven v4 Turbo?

Eleven v4 is tuned for quality and long-form expressive reads, up to 10,000 characters per request. v4 Turbo is tuned for real-time agents, with a claimed ~100 ms median inference latency.

Does Eleven v4 have open weights?

No. It’s available only through ElevenLabs’ apps and API. The highest-ranked open-weights model on the Artificial Analysis arena is Breeze TTS 2, which is non-commercial.

Accurate as of October 2, 2026. Pricing, promotions and leaderboard positions move fast. More free AI voice and creative-tool guides are on Creative AI Show.