Voice Agent · HeyGen × OpenAI
The avatar just learned
to cut you off mid-sentence.
HeyGen's video avatar, LiveAvatar, now runs on OpenAI's new voice model, GPT-Live-1. Both companies open-sourced the full implementation of a voice that listens while it talks and reacts to interruptions in real time. Here's where an avatar that used to just read a script from top to bottom is headed next.
The Release
Two announcements,
two days apart
First, on September 10, 2026, OpenAI shipped a new API model called GPT-Live-1. Its defining trait is full-duplex voice: it listens to the other party while it's speaking and can react to backchannels or interruptions on the spot. It's priced at $0.05 per minute, ships with 12 voice options, supports telephony, and follows instructions more reliably than earlier voice models.
The very next day, September 11, video-avatar company HeyGen announced it had integrated GPT-Live-1 into LiveAvatar. HeyGen's official X account posted that "real-time AI experiences are now solved AND open sourced," describing a close collaboration with OpenAI to combine GPT-Live-1, LiveAvatar, and HyperFrames into a single framework — and published the working implementation as a repository. This wasn't just a partnership announcement — it was a release of runnable, working code, and that's what makes it worth a closer look.
The published repository, heygen-com/liveavatar-gpt-live-demos, ships under the MIT license (with bundled GSAP under its own separate license). The default demo is a Japanese-language tutor: it speaks vocabulary out loud, puts a card with the term, pronunciation, and meaning on screen as it does, and every few words shrinks itself into a corner to review what's been taught so far.
How it's actually wired
The "voice" and the "brain" of the avatar are split into two separate models.
Per the repo's own implementation notes, audio and control data travel on separate paths. Mic audio goes up to GPT-Live-1 (internally the "GPT-Live v3" contract, model id gpt-live-1) over a WebSocket, while transcripts and on-screen display instructions come back down the same channel. The raw audio itself is routed to a media server without passing through the browser, and the authors say they deliberately avoided framework-style abstraction so the bare implementation can be forked and run as-is.
Who It Changes
Who feels this, and how
Engineers
An MIT-licensed reference implementation is now sitting there for the taking, which sharply cuts the cost of building this into your own product. With Node 20.12+, pnpm, and API keys for both services, you can fork it and start modifying today.
Designers / Marketers
You can now prototype an avatar that actually holds a conversation — for a sales demo or live stream — instead of one that reads from a script. That said, the official demo tops out at 12 voices and a Japanese tutor use case, so tailoring it to your own scenario still takes real work.
Script-first production teams
For work that's fully served by reading a script — like a straightforward promo video — full-duplex conversation buys you almost nothing. There's little reason to change an existing workflow.
Next / Risks
What comes next, and where it snags
What's next — now that the implementation is out under MIT, it's only a matter of time before other avatar vendors and in-house teams port it. The $0.05/minute price point puts the real question front and center: does the math work for an always-on, long-duration front desk or call-center-style use case? Start by running the numbers on your own expected call volume and comparing them against your current voice-response system.
The counterpoint / risk — based on the repository's own description, the voice itself is handled by GPT-Live-1, while the decision of what to show on screen is delegated to a separate backend model. That means the more natural the conversation feels, the harder it becomes for a user to notice when the backend gets something wrong — a confidently delivered wrong answer sounds just as fluent as a correct one. Anyone deploying this for customer-facing or educational use needs a fact-checking layer that sits outside the conversation itself.