The 12 Best Real-Time AI Avatar APIs in 2026, Compared on Price, Latency and Integrations

• Decide first between a face layer and a managed agent. A face layer animates your agent's audio; a managed agent also runs speech, the LLM and the voice. Their per-minute prices are not comparable.
• At 2,000 minutes a month, face layers cost from $19 (Spatius) and $20 (Simli) to $342 (LemonSlice); managed agents from $120 (Voice Agents, ours) to $614 (Tavus). D-ID's self-serve plans stop at 600 minutes.
• Integrations often decide the pick. LiveKit lists 16 avatar plugins, and ElevenLabs documents one avatar integration for its Agents: HeyGen LiveAvatar.
• Concurrency and call caps are the hidden limits. HeyGen LiveAvatar's paid plans and Talking Avatar have no concurrency cap; Runway ends every session at 5 minutes.
• Older lists are out of date. HeyGen's Interactive Avatar now lives on as LiveAvatar, and Hedra's real-time avatar service has shut down.

These are our picks by job, not a single ranking. Each API wins somewhere.

A face layer (vendors call it LITE mode, speech-to-video, audio passthrough or an avatar plugin) takes the audio your voice agent already produces and returns lip-synced video. You keep your own speech recognition, LLM and voice, and pay for them separately. HeyGen LiveAvatar LITE, Anam, Simli, Beyond Presence's Speech-to-Video API, LemonSlice, Spatius, bitHuman, Synthesia and Talking Avatar Realistic Avatars all work this way.

A managed agent runs the whole conversation: it listens, thinks with an LLM, speaks and renders the face. Tavus CVI, D-ID Agents, HeyGen LiveAvatar FULL, Beyond Presence Managed Agents, Runway Characters and Talking Avatar Voice Agents work this way. A managed minute costs more because it includes the models.

Comparing a $0.02 face layer with a $0.24 managed agent is a common budgeting mistake. Add your own LLM and voice bill to every face-layer price before you compare it with a managed one. Our AI voice agent pricing guide shows what that bill usually adds.

We priced 2,000 conversation minutes a month, about 65 minutes a day, on the cheapest self-serve plan that covers them, plus overage at each vendor's published rate. Prices are as of 30 October 2026.

Plan limits matter as much as rates. Anam's Explorer plan ends calls at 10 minutes, so longer calls need Growth at $299. Spatius's $19 Starter plan ends calls at 30 minutes and allows 4 sessions at once, and Simli's Hobby plan allows 2.

The crossover points are simple to check. Leaving aside Simli's free 50 minutes a month, Talking Avatar Realistic Avatars cost less than Simli below 500 minutes a month, because Simli's $10 Hobby plan buys 500 of our minutes. They also cost less when you need more sessions at once than a Simli plan allows. Simli Pro ($49, 10 sessions) equals 2,450 of our minutes, and Simli Scale ($249, 50 sessions) equals 12,450. Above those volumes, with steady traffic, Simli costs less. Against HeyGen LiveAvatar LITE, once you pass its free 10 credits a month, ours cost less at every volume on published self-serve plans: $0.02 a minute against LiveAvatar's lowest rate of $0.079 (Business plan). LiveAvatar wins elsewhere: sharper video (720p by default, 1080p optional, against our 832×468), LiveKit, Pipecat and ElevenLabs integrations, and a FULL mode that runs the conversation for you. For other volumes, see our real-time avatar API pricing table of 25 providers and the cheapest AI avatar APIs.

Every vendor says “real-time”, but they measure different events, so treat these as each vendor's own claims, not a benchmark:

• Simli publishes a response time of about 150 ms on its homepage.
• Beyond Presence quotes 250 ms from audio input to avatar video in its voice agent docs.
• Talking Avatar: less than 250 ms, our own figure. Measured on our servers across production sessions (October 2026): the time from your agent's audio reaching the avatar to the avatar speaking it. The network between our servers and your caller adds to what they hear.
• HeyGen LiveAvatar reports a median time to first frame under 300 ms on its site.
• LemonSlice reports a 471 ms average time to first byte for its Flash model, from its own benchmark.

The avatar adds only part of the delay. Your speech recognition, LLM and voice usually add more, so test each candidate with the same agent, voice and network before you choose.

LiveKit's avatar documentation lists 16 plugins as of 6 November 2026: Anam, Beyond Presence, D-ID, LemonSlice, Protoface, Runway, Tavus and TruGen in Python and Node.js, and Avatario, AvatarTalk, bitHuman, Keyframe, LiveAvatar, Simli, Spatius and Synthesia in Python only. Pipecat's documentation lists video avatar services for Anam, bitHuman, HeyGen, Protoface, Simli and Tavus, and transports for Beyond Presence, HeyGen, LemonSlice and Tavus.

If you run LiveKit Agents

Pick a listed plugin and the avatar becomes a second participant in your room, publishing synced audio and video. Anam, Beyond Presence and Tavus are the strongest picks with both Python and Node.js support. Talking Avatar Realistic Avatars join a LiveKit room the same way: on Python with our own plugin, livekit-plugins-talkingavatar on PyPI (in review for LiveKit's list), and on Node.js through our SDK, with a publish-only token either way.

If you run ElevenLabs Agents

HeyGen LiveAvatar is the documented route: ElevenLabs' own guide connects an agent to LiveAvatar's LITE mode, billed at 1 LiveAvatar credit a minute plus your ElevenLabs usage. D-ID and Anam publish their own ElevenLabs guides. Talking Avatar works with ElevenLabs text-to-speech, but we don't document an ElevenLabs Agents integration.

If you run the OpenAI Realtime API

Any face layer that accepts raw audio works if your Realtime session runs on your server over WebSocket. Our docs pass Realtime's 24 kHz PCM audio deltas straight to the avatar. If your page talks to OpenAI Realtime directly over WebRTC, the audio never reaches your server, so move the session server-side first.

If you run a phone agent on Vapi or Retell

Callers on a phone line can't see a face. An avatar only helps when the conversation moves to a website or app, where you can use any face layer above or a no-code web agent. For phone agents, see our 12 best AI voice agent platforms.

The numbers are for reference, not a ranking; the picks by job give the best API for each use.

Also considered

• Akool Streaming Avatar costs about $0.16 to $0.35 a minute at 1080p, with API access from its $59 Pro Max plan and 3 to 10 sessions at once (Akool API pricing).
• Google's Gemini Live API can return an avatar for about $0.37 per speaking minute of video, billed only while the avatar speaks, plus voice and model tokens (Vertex AI pricing); custom faces are for select customers (Google Cloud docs).
• Microsoft Azure TTS Avatar costs $0.50 to $0.80 a minute plus Azure text-to-speech, at 1920×1080 and 25 fps, with Azure voices only (Azure pricing).
• Smaller face layers listed as LiveKit plugins run from $0.01 to $0.17 a minute: Protoface ($0.01), Avatario ($0.05), AvatarTalk ($0.05 to $0.10) and Keyframe Labs ($0.10 to $0.17).
• UneeQ and Soul Machines sell enterprise digital humans; AppDirect announced it had acquired Soul Machines on 28 September 2026 (SiliconANGLE).
• Hedra is left out: its real-time avatar service has shut down, and LiveKit's Hedra plugin says the service is disabled.

Coming from HeyGen's Interactive Avatar? See our HeyGen Interactive Avatar alternatives and Tavus alternatives.

The EU AI Act is Regulation (EU) 2024/1689.