Rabbit sold a hundred thousand people a $199 AI box to run their apps, then most of them stopped using it within weeks.

The pitch — name stripped

What actually happened

Rabbit launched into one of the loudest hype waves in recent gadget history. Founder Jesse Lyu, a two-time Y Combinator alum who had sold an earlier conversational-AI startup to Baidu, raised around $30 million from Khosla Ventures and others, and unveiled a small orange box designed by Teenage Engineering at CES in January 2024. The pitch was a 'Large Action Model' — an AI that, unlike a chatbot, would take real actions in your apps, moving AI 'from words to action.' It went viral: roughly $10 million in pre-orders and more than 100,000 units, with the first batches selling out in days.

Then the R1 shipped, and it did not work. Reviewers found the Large Action Model slow and unreliable: ask it for an Uber and it got the pickup and drop-off wrong; ask it to play a song and it would acknowledge the command and then fail to play anything. Battery life was poor, the software was buggy, and the grand 'it can do anything' promise collapsed into a handful of half-working integrations. The device did less, worse, than the phone already in the buyer's pocket.

Then came the question that never went away. An Android developer showed that the entire R1 experience ran as a single Android app on a stock open-source Android build — installable and runnable on an ordinary Pixel phone. Whatever the model was doing in the cloud, the thing customers paid $199 for looked a great deal like software that needed no special device at all. The founder disputed it, but the doubt stuck: if it's an app, why is it a gadget?

The usage numbers settled it. By September 2024, The Verge reported that only about 5,000 of roughly 100,000 buyers were using the R1 at any given moment. The pre-orders had proven the design and the story could sell; daily use proved the product could not hold anyone. A device people stop carrying isn't a platform, it's a souvenir — and the R1 became one within weeks.

Agentic AI — software that takes real actions on your behalf — is a genuine and coming thing, and Rabbit was early to name it. But it shipped the promise as a separate $199 box before the agent could reliably do the jobs, and before there was any reason the box had to exist instead of an app running on the phone that already drives all those same services.

The lesson

Kasspian's read went where it goes on every overbuilt gadget: the riskiest assumption was never 'can we make a charming AI box,' it was 'can an AI agent reliably run apps built for humans, and does it even need its own hardware to do it.' Both halves were unproven, and the launch demo glossed over both. A polished reveal film is not a working product, and pre-order volume measures your story, not your software.

'A new device to replace the phone' is the trap again — but Rabbit stacked a second unproven bet on top: the AI agent itself barely worked. When the core capability is the part you haven't proven, putting it inside expensive hardware just raises the stakes on a bet you haven't won. And when a developer can run your whole product as an app on a phone, the device was never the point — the device was the marketing.

The cheap test was the agent, not the gadget. Before the tooling, the orange box, and the CES stage, the question was simple: can this thing book a ride and play a song correctly, every time, for a stranger? Ship that as an app, prove people use it daily, and only then ask whether it deserves its own hardware. Pre-orders measure how good your story is; daily use measures whether you built anything. Rabbit mistook the first for the second.

Common questions