Questions worth a straight answer.

Free because it barely costs us anything. Your phone draws every frame and speaks every word; the only thing running on our hardware is the conversation model, and a small model on our own GPUs works out to a fraction of a cent per hour of chat. Vendors who ration free tiers do it because every message costs them real inference. That pressure doesn’t exist here.

The Companion is time-capped, not metered — an hour a day free, uncapped on Plus. If you use a frontier cloud model or a cloud voice, those are billed per minute because they genuinely cost us per minute, and you see the price before you turn them on. The included brain and the on-device voice stay free.

One of them renders on-device and still bills a per-minute rate through a heartbeat their SDK posts every sixty seconds. Others stream compact pose data and rasterise it in the browser — still a server generating motion for every frame of every session. Turning audio into motion is the expensive part, and here it happens on the phone.

Rendering and speech do — no audio, video, frames or transcripts leave the device. The conversation model does need a link, because it runs on our GPUs. In the SDK you bring your own brain, so what crosses the network is entirely your choice.

It runs on the Neural Engine rather than the GPU, which is what makes long sessions viable — much less power, and none of the thermal cliff a GPU hits under sustained load. We have not published a measured curve yet and we are not going to invent one. When we do it will be against the honest baseline: a streamed avatar spends the same battery on continuous video decode and a busy radio.

No — the Companion is licensed for personal, non-commercial use. Anything that ships inside a product you sell, or that puts a Yoob character in front of your customers, needs the SDK. If you are a student, researcher or hobbyist and unsure which side of that line you are on, email us; we will almost certainly say yes.

Eventually, yes, and we will not tell you otherwise — a determined attacker can pull a Core ML model out of any iOS app. Weights are never in your binary, they arrive over an attested channel encrypted at rest, and each licensee’s copy is fingerprinted so a leak identifies its source. Security has the detail.

Because a cloud vendor can afford a large model on a large machine, and we have to land an identity inside something that runs at 25 fps on a phone. Training, quantisation, Core ML conversion and a full pose-range check take days. Stock characters ship the same day.

iOS ships today. Android is in development, Web is on the roadmap, and neither has a date we are willing to print. If your timeline depends on one, tell us — it moves our priorities.