Spatius vs Talking Avatar: Price, On-Device Rendering and Limits

As of 30 October 2026, Spatius sells monthly plans: Starter at $19 for 2,000 minutes, Builder at $49 for 5,500, Growth at $149 for 18,000 and Scale at $299 for 40,000. That is $0.0095 to $0.0075 a minute. Annual billing saves 20%: Starter is $189 a year. A minute uses 10 credits, and paid plans keep running past their minutes at $0.009 a minute ($0.007 on Scale), or $0.0072 and $0.0056 on annual plans. Spatius advertises "$0.42 per hour", the Scale rate for extra minutes; ours works out to $1.20 an hour. Enterprise starts at $699 a month for 80 sessions at once and 80,000 minutes, then $0.005 a minute, and never costs more than $5,200 a month. The Free plan's 100 minutes a month exclude commercial use. Spatius marks its pricing as beta: "pricing and limits may change".

At steady volume, almost always. $19 buys 950 minutes from us, so past about 950 minutes a month Spatius Starter costs less. 10,000 minutes cost $200 on Talking Avatar Realistic Avatars and $89.50 on Spatius Builder ($49 plus 4,500 minutes at $0.009), or about $71 billed annually. We cost less in three cases. Light commercial use: under about 950 minutes a month, our per-minute price beats any subscription. Peaks: Starter allows 4 sessions at once, Builder 8, Growth 22 and Scale 40, so a class of 15 students practicing at the same time needs Growth at $149 even for 2,000 minutes, which cost $40 on our side. Long calls: Starter ends sessions at 30 minutes, so 45-minute interviews need Builder and 90-minute tutoring needs Growth; ours run up to 3 hours. At very high volume both sides quote: our enterprise and high-volume contracts go as low as $0.005 a minute, and Spatius Enterprise charges $0.005.

Spatius doesn't stream video. Its Motion Server turns your agent's speech into motion data, about 10–15 KB a second, and AvatarKit draws a 3D Gaussian-splatting avatar on the user's device, with WebGPU in Chrome and Edge and WebGL elsewhere. That is why it costs so little and uses so little bandwidth. The device pays instead: it downloads the avatar, about 5–10 MB, before the first frame (later loads can come from cache), and renders every frame. The SDK includes a benchmark, isDeviceSupported(), to check whether a device can run it, and web builds need WebAssembly set up in Vite or Next.js. Talking Avatar Realistic Avatars renders on our servers and sends video over WebRTC, so the viewer's device only plays a video stream, like any video call, and nothing ships in your app bundle. Our avatar starts speaking less than 250 ms after your agent's audio reaches our servers; Spatius publishes no comparable figure.

Spatius has a LiveKit Agents plugin, livekit-plugins-spatius, for Python; LiveKit's docs list no Node.js version. The plugin sends your agent's speech to Spatius and publishes audio and motion data into the room. LiveKit's own docs explain the catch: the avatar's video track carries motion data in otherwise black frames, so a standard video renderer, including the Agent Console and the LiveKit Playground, shows a black screen. Your frontend has to run Spatius's SDK, and its packaged LiveKit client is web only; on iOS and Android you supply the LiveKit provider yourself. Talking Avatar Realistic Avatars joins your room as an ordinary participant with a video track and a voice track, so any LiveKit client shows it, on the web, iOS, Android, Flutter or React Native. You start it from our Node.js SDK with a publish-only token you mint, or from Python or any other language through our realtime WebSocket API.

Spatius avatars come from its public library or from one image in Spatius Studio. Paid plans include 2, 4, 14 or 40 avatar creations a month, and extra creations cost $25 each, down to $16 in a pack of 500. AvatarKit renders the avatar on a transparent background, so it can sit over your own image, video or 3D scene. Spatius runs in US West, Tokyo and Beijing; Enterprise lets you choose the region and adds 99.5% availability with service credits and a DPA on request. Talking Avatar Realistic Avatars offers a growing library of ready-made photorealistic avatars, and a custom avatar of your own presenter costs $199 once, made from a short video and voice recording with their consent, and it works in the API and in Voice Agents. Spatius's FAQ says it doesn't provide a voice agent. We sell two ways in: Realistic Avatars for the agent you already run, and Talking Avatar Voice Agents, built in our dashboard, at $0.06 a minute with the LLM, speech recognition and voice included.

Spatius and Realistic Avatars both take your agent's speech, so switching changes how the face reaches the viewer and keeps the agent. The bigger change is on the client: AvatarKit comes out.

1. Create a Talking Avatar Realistic Avatars account (it comes with $1 of free credits), create an API key and pick an avatar ID from the library, or order a custom avatar for $199.
2. In your LiveKit worker, remove spatius.AvatarSession and keep the agent's speech recognition, LLM and TTS.
3. Mint a publish-only LiveKit token for the avatar and start the session with it, from our Node.js SDK or, for a Python worker, through our realtime API (our docs have a complete Python example).
4. Send your agent's TTS audio to the session as PCM16 at its native rate (8 to 48 kHz). The avatar publishes the voice in sync with its lips, so don't also play the agent's own voice track.
5. In your frontend, replace AvatarKit and its WebAssembly build setup with a plain LiveKit video element, or our browser component if you don't use LiveKit.
6. Run both on the same agent for a week and compare cost and how the avatar looks on your users' slowest devices before you cancel the Spatius plan.