Search for “live avatar AI” and you get the rock band LIVE, a dictionary entry for the word “live,” and a stack of vendor pages — each describing its own product as if it were the whole category. The developers behind that search are usually after something narrower and more concrete: a real-time conversational avatar they can put behind an API, and a price they can put in a spreadsheet.
Live avatar AI is not one product, and the word “live” carries most of the confusion. This guide separates the three things vendors actually ship — scripted avatar video, streamable batch generation, and real-time conversational avatars — then normalizes pricing for the vendors that rank for the term (HeyGen, D-ID, JAI, New Port AI) to a per-minute basis where the published data allows it, walks through the documented integration shapes, and lists the failure modes that burn money. Every figure is date-stamped, accessed August 27–28, 2026. Full disclosure: Modellix is an AI model API aggregator with a commercial interest in the routes it sells, so we state plainly where Modellix fits in this market — and where it does not.
What counts as “live” in live avatar AI
“Live” is used loosely across the whole category. Before comparing anything, separate the three families — they have different latency budgets, different architectures, and different price points:
- Scripted avatar video. You type a script, the platform renders a video of an avatar speaking it (HeyGen’s studio, Synthesia, most of D-ID’s non-realtime output). Nothing is live: the render takes minutes, and the result is a video file you host. This is what most people mean by “talking avatar,” and our talking avatar API guide covers the API side of it in depth.
- Streamable batch generation labeled “live.” A model animates a portrait from an audio file and streams the result as long-form video. JAI’s Live Avatar is the cleanest example: real-time animation from a single reference image plus an audio track, streamable in near-“infinite-length” clips of roughly three seconds each. It is fast and streamable, but it is not conversational — there is no listening loop, no STT, no turn detection. The lip sync is the point.
- Real-time conversational avatars. The avatar listens, reasons, and speaks in a loop over WebRTC: speech-to-text, turn detection, an LLM response, text-to-speech, and synchronized avatar rendering, streamed back to the user in near real time. This is what HeyGen LiveAvatar, D-ID’s realtime agents, and LiveKit-orchestrated stacks deliver, and it is the only family where “live” is architecturally honest. D-ID’s own glossary defines a real-time avatar exactly this way: conversational AI delivered through an avatar over WebRTC.
Most of the SERP for “live avatar AI” fails to draw this line, which is why the same product gets quoted as both “sub-500 ms” and “not really live” depending on the page. When a vendor page says “live,” ask which family it belongs to before you evaluate latency or price. A streamable-batch model at $0.02/minute-equivalent cannot compete with a conversational avatar that needs an STT→LLM→TTS round trip on every turn — and it does not have to; they are different products.
The three families that ship under “live avatar AI”: scripted video, streamable batch labeled live, and real-time conversational over WebRTC. Illustrative concept diagram.
The real-time avatar pipeline: what you’re actually integrating
A conversational avatar is a voice pipeline with a video output stage. D-ID’s realtime documentation describes the canonical shape: user speech flows through speech-to-text, then turn detection (deciding when the user has finished speaking), then the LLM (which can query an optional knowledge base for RAG answers), then text-to-speech, then the avatar renderer, and the video is delivered back to the user over WebRTC. LiveKit’s avatar documentation describes the same delivery layer from the orchestration side: an avatar worker joins the room as a separate participant and publishes synchronized audio and video tracks.
The design decision that actually shapes your integration is who manages which stage:
- Vendor-managed full mode. HeyGen’s FULL mode runs the whole pipeline — WebRTC, STT, LLM, TTS — and you pay 2 credits per minute for the privilege. You integrate by requesting an embed and dropping it into a page. This is the fastest path to a working avatar.
- Bring-your-own stack. HeyGen’s Avatar Only (LITE) mode charges 1 credit per minute and hands you the WebRTC, STT, LLM, and TTS stages. D-ID’s realtime agents support client keys for a backendless embed, a backend session token for server-controlled sessions, or a full SDK for a custom UI. LiveKit’s avatar plugins — for Anam, D-ID, LemonSlice, HeyGen LiveAvatar, Tavus, and others — give you the same split inside an existing agent framework.
Your choice between the two is a latency and cost trade, not a quality judgment. If your product already has an LLM and a TTS provider you like, LITE-mode or a LiveKit plugin avoids paying for pipeline stages you would otherwise run twice. If you need an avatar in a demo by Friday, FULL mode wins — 2 credits per minute is the price of not operating the pipeline yourself.
Azure’s text-to-speech avatar documentation is worth a read even if you never touch Azure: it is one of the clearest public explanations of how real-time synthesis differs from batch avatar generation, which is the same distinction this article is drawing.
The real-time avatar pipeline every conversational product shares: STT → turn detection → LLM → TTS → avatar, delivered over WebRTC, with an optional knowledge base on the LLM. Illustrative concept diagram.
Live avatar pricing, normalized to per-minute
Every vendor bills in a different unit — credits per minute, monthly minute allowances, or pay-as-you-go clip credits — so the only honest comparison is to normalize to a per-minute cost where the vendor publishes enough to do so. All figures below were read from the live pages on August 27–28, 2026. Where the data is insufficient, this table says so instead of inventing a conversion.
| Vendor | Published plan / unit | Published allowance | Effective per-minute |
|---|---|---|---|
| HeyGen (LiveAvatar FULL) | Creator $29/mo, 600 credits | FULL mode = 2 credits/min | ≈ $0.097/min (calculation from list prices) |
| HeyGen (Avatar Only / LITE) | same plans | LITE mode = 1 credit/min | ≈ half of FULL (same calculation) |
| D-ID | Lite $4.7/mo, Pro $16/mo, Advanced $108/mo | 10 / 15 / 100 min per month | not published per minute for realtime agents |
| JAI Live Avatar | pay-as-you-go credits | ~3-second clips; frames_per_clip/acceleration change cost |
no per-minute list price on the accessed page |
| New Port AI | Professional $49/mo, Premium $259/mo | 300 / 2,000 min per month | ≈ $0.163/min and ≈ $0.13/min (calculation) |
A few notes on what is — and is not — in that table:
- HeyGen publishes its plan prices (Free $0 with 3 videos/month, Creator $29, Pro $49, Business $149) and its LiveAvatar credit rates (FULL 2 credits/min, Avatar Only 1 credit/min) in separate places; combining them is our arithmetic, not an official rate. $29 ÷ 600 credits × 2 credits/min ≈ $0.097/min at Creator list prices — before any credit rollover or volume discounts.
- D-ID publishes its studio/API plan prices (Lite $4.7/mo for 10 minutes, Pro $16/mo for 15, Advanced $108/mo for 100) and rounds minutes up to the nearest 15-second interval, but it does not publish a per-minute rate for realtime agents on the pages we accessed — so we do not compute one. Realtime agents sit in the higher tiers.
- JAI sells Live Avatar on credits with no per-minute list price on the accessed page; cost depends on clip count, frames per clip, and acceleration. Its “10 free credits — no credit card” trial is the practical way to measure your own per-minute cost.
- New Port AI publishes monthly minute allowances (Free 30 min/mo, Professional $49/mo for 300, Premium $259/mo for 2,000); the per-minute figures are our division of published numbers.
One guardrail before you build a budget: credits ≠ dollars ≠ minutes. Never divide by a conversion you invented. If a vendor does not publish a per-minute rate, the honest move is to run a timed trial and meter it yourself — most of them fund one (HeyGen’s free tier, D-ID’s 14-day trial, JAI’s 10 credits, New Port AI’s 30 free minutes).
A working live avatar integration
The request shapes below follow each vendor’s documented contract — they are written from the docs, not executed against a live account, so treat field names as current-but-verify. All three paths are realistic integration points.
HeyGen — fastest path (FULL mode embed). The quickstart in HeyGen’s LiveAvatar docs is a single call: POST /v2/embeddings with your API key returns a short-lived embed URL and a ready-to-use script tag.
1 | curl -X POST https://api.heygen.com/v2/embeddings \ |
1 | { |
Drop the returned script value straight into your HTML — it is a ready-to-use iframe. If you need your own UI or custom conversation logic instead, HeyGen publishes a Web SDK and supports LiveKit or Agora for custom WebRTC handling.
D-ID — client keys or backend session token. D-ID’s realtime path lets you embed a prebuilt agent widget with client keys and no backend at all; for server-controlled sessions you create a session on your backend, obtain a token, and pass it to the client, which uses it to connect to the stream. The talk stream reference documents the stream contract if you want to drive the WebRTC connection yourself. The realtime overview is the map of which component (STT, LLM, TTS) is optional and which provider plugs in where.
LiveKit — avatar session inside an agent. With a LiveKit avatar plugin, you add an avatar to an existing voice agent in a few steps: install the plugin and API keys, create an AgentSession, create an AvatarSession configured for your provider, start the avatar session, then start the agent session with audio output disabled — the audio goes to the avatar worker, which publishes synchronized audio + video to the room. Call wait_for_join() before session.start() so the avatar video is live before the agent starts speaking; it waits up to 30 seconds by default and raises asyncio.TimeoutError if the avatar never joins. In the frontend, distinguish the agent from the avatar worker participant (kind agent with publish_on_behalf set) so you render the worker’s tracks, not the agent’s.
Which one is “the” integration depends entirely on your stack: no backend wanted, an existing agent framework, or a full custom pipeline. The pricing section is the tiebreaker — FULL mode is the most expensive per minute precisely because it removes the most work.
Avatar & Video Model API Reference
See the REST contract for Modellix's avatar-adjacent routes — kling-avatar, single-actor-avatar and 200+ image and video models on one key.
View DocsFailure modes that cost money
Integration guides skip the failure modes because they are boring — but for live avatars, failures are metered. The expensive ones:
- Credits burn during FULL mode. At 2 credits/min in HeyGen FULL mode, an idle session that stays connected while a user thinks burns the same credits as one where they talk. If your conversation has long silences and you do not need HeyGen’s pipeline, LITE mode halves the rate — or you manage the session lifecycle yourself.
- WebRTC join and playback latency. LiveKit’s
AvatarMetricsreports two numbers worth alerting on: join latency (time from avatar session start until the avatar publishes video) and playback latency (delay between the agent forwarding audio and playback starting on the avatar). A high playback latency means the avatar starts moving its mouth late — which users read as broken. - Minutes rounding. D-ID rounds usage up to the nearest 15-second interval per its pricing FAQ. Short sessions round up, and at scale the rounding is not noise: 100 sessions averaging 8 seconds bill as 100 × 15 seconds.
- Content-policy rejections mid-flight. Avatar providers run moderation on generated speech and rendering; a rejected turn can fail the session after you have already paid for connection time. JAI’s model page advertises an integrated safety checker; D-ID’s terms require review before production use of certain voices/avatars.
- Audio-video desync. This is the failure users notice first. It is usually a pipeline-latency problem (LLM slow to respond → TTS late → avatar mouth out of sync with intent), not a rendering bug. Instrument the per-stage latencies before you blame the avatar vendor.
- Regional STT/TTS latency. The LLM can be 50 ms away and the STT 800 ms away if they are in different regions. Co-locate the pipeline or the round-trip budget quietly doubles.
The pattern behind all of these: measure before you meter. Every vendor charges by wall-clock time or credits, and every failure mode above inflates one of those two numbers.
Where Modellix fits — and where it does not
Honest catalog statement first: as of August 27, 2026, Modellix does not list a real-time streaming avatar API. Site search for live-avatar, realtime-avatar, and streaming-avatar returns zero hits, and neither the AI avatar generator collection nor the digital-human routes include a WebRTC conversational product. For a true live-streaming avatar, you cannot go through Modellix at all. That is not a claim that the aggregator route is cheaper — for this category there is no aggregator route, and we are not going to pretend otherwise.
What Modellix does carry is the avatar-adjacent pipeline: scripted avatar generation models like kling-avatar (image-to-video avatar, $0.0448–$0.0896/sec on the model page, August 28, 2026) and skywork single-actor-avatar (audio-to-video talking head from one portrait and one audio track, $0.0480–$0.0720/sec), video lip-sync models covered in our pixverse lip-sync guide, and TTS models (qwen-audio and the CosyVoice family) that can generate the audio side of an avatar pipeline. Those are family 1 and family 2 products — scripted and streamable batch — not family 3.
That limitation is also the workflow argument for the adjacent use case. If your product is “generate avatar videos from a portrait and a script,” you can build that pipeline entirely through one key and one submit–poll contract: how the Modellix unified API works explains the shared contract, and the AI avatar API overview walks the catalog. If your product needs a conversational avatar, go direct to HeyGen, D-ID, or a LiveKit-orchestrated provider — and re-check this page when the aggregator catalog grows, because a realtime avatar route is exactly the kind of model that would land here.
How to choose a live avatar API
The decision framework is simpler than the marketing suggests. Match your latency budget to a family, then let pricing and pipeline ownership decide the vendor.
| If you need… | Start with… |
|---|---|
| Sub-second conversational latency, turn-based dialogue | HeyGen LIVE / Avatar Only, D-ID realtime agents, or LiveKit + a provider plugin |
| Minutes-latency streamable video from a portrait + audio | JAI-style streaming models (streamable batch, not conversational) |
| Offline scripted avatar video at scale | HeyGen studio, Synthesia, or Modellix’s scripted avatar routes |
| Several avatar models behind one key and one bill | Modellix’s avatar-adjacent catalog (scripted generation, lip sync, TTS) |
The free-tier reality check: HeyGen’s Free plan includes 3 videos/month, D-ID runs a 14-day trial with 3 minutes, JAI gives 10 free credits, and New Port AI includes 30 minutes/month free. None of those funds a production pilot by itself — but together they let you time the same conversation script on each platform and measure real per-minute cost and real lip-sync latency, which is worth more than any vendor’s spec sheet. (For the budget-constrained path, our free talking avatar options roundup covers the no-cost routes in detail.)
One compliance note: if your use case is customer-facing and regulated, check what each vendor publishes before you commit. D-ID documents SOC 2 and ISO 27001/27017/27018/27701 certifications on its trust pages, which is more than most avatar vendors disclose; ask for the equivalent from anyone else before you send real customer audio through them.
Explore Avatar Models on Modellix
Log in to check live pricing on kling-avatar, single-actor-avatar and 200+ image and video models behind one API key.
LoginFrequently Asked Questions
What is live avatar AI?
Live avatar AI refers to avatars that render and speak in real time, usually as part of a conversational loop — the avatar listens, an LLM formulates a response, TTS speaks it, and the avatar’s lips and gestures are rendered in sync over WebRTC. The term is also loosely applied to streamable batch generation that animates a portrait from an audio file without a conversation loop.
Is HeyGen LiveAvatar an API?
Yes. HeyGen exposes LiveAvatar through POST /v2/embeddings, which returns a short-lived embed URL and a script tag, plus a Web SDK and LiveKit/Agora options for custom WebRTC handling. FULL mode (2 credits/min) runs the pipeline for you; Avatar Only / LITE mode (1 credit/min) leaves STT, LLM, and TTS to you.
How much does a live avatar AI cost per minute?
It depends on who runs your pipeline. HeyGen’s published rates work out to roughly $0.097/min in FULL mode and about half that in Avatar Only mode at Creator list prices (our calculation, not an official rate). New Port AI’s plans divide to about $0.163/min (Professional) and $0.13/min (Premium). D-ID and JAI do not publish per-minute realtime rates on their accessed pages, so an honest answer for them requires a timed trial.
What is the difference between a live avatar and a talking avatar?
A talking avatar is any avatar that speaks — including scripted video rendered in minutes. A live avatar is specifically one that renders and responds in real time within a conversational loop. Every live avatar is a talking avatar; most talking avatars are not live.
What latency can I expect from a live avatar API?
Real-time conversational avatars are engineered for sub-second to ~2-second turn times (vendors quote under 500 ms lip-sync to under 2 s conversational timing), but the number you get depends on your own pipeline: STT, LLM, and TTS each add latency, and misconfigured regions can quietly double the round trip. LiveKit’s AvatarMetrics (join latency, playback latency) is the standard way to measure the delivery stage.
Is there a free live avatar AI?
Every major vendor has an entry tier: HeyGen’s Free plan (3 videos/month), D-ID’s 14-day trial (3 minutes), JAI’s 10 free credits, and New Port AI’s 30 minutes/month. These are evaluation allowances, not production free tiers — and none of them is a free realtime API with unlimited minutes.
Does Modellix offer a live streaming avatar API?
Not as of August 27, 2026. Modellix does not list a real-time streaming avatar API; its catalog covers avatar-adjacent routes — scripted avatar generation (kling-avatar, single-actor-avatar), video lip-sync, and TTS — through one key and one submit–poll contract. For a true live-streaming avatar, use HeyGen, D-ID, or a LiveKit-orchestrated provider directly.
Pricing and availability sources accessed August 27–28, 2026: HeyGen LiveAvatar docs and HeyGen pricing, D-ID realtime overview, D-ID talk stream reference, D-ID pricing, D-ID real-time avatar glossary, LiveKit avatar docs, HeyGen LiveAvatar Web SDK, JAI Live Avatar, Azure text-to-speech avatar docs, and New Port AI. Provider pricing and availability change frequently; validate against each live page before committing budget. Modellix is an AI model API aggregator with a commercial interest in the routes it sells, and this guide is not a claim that the aggregator route is cheaper — for a true live-streaming avatar you currently cannot go through Modellix at all.
Cover image is illustrative Modellix artwork, not a provider screenshot or pricing evidence.
