Rabbit R1: The $199 device that wants to use your apps

On January 9, 2024, Rabbit introduced R1 at CES as a $199 device intended to operate digital services through natural-language requests. The original announcement specifies price, pre-order, expected shipment and the Large Action Model promise; a prepared demonstration does not yet prove that an everyday transaction will complete reliably.

Designed with Swedish firm Teenage Engineering, the device stands out with its orange casing and compact form factor. But it is not intended to compete with the smartphone as another phone: Rabbit presents the R1 as a physical controller for digital services. The company says the first units available for preorder have sold out.

An assistant designed to act, not just respond

The R1 runs Rabbit OS, a proprietary operating system based on what the company calls a Large Action Model, or LAM. The term recalls the large language models behind chatbots such as ChatGPT, but Rabbit is trying to solve a different problem.

A language model generates or interprets text. Rabbit’s action model is designed to understand the steps required within an interface and complete them: search for a song, request a ride, order food or manage a purchase. Rather than simply recommend an app or display instructions, the device promises to carry out the task.

The distinction matters. Voice assistants have been able for years to answer simple questions, set alarms and control smart-home devices. But they typically rely on specific integrations with each service. Rabbit says its model can learn to use apps as a person would, navigating their screens.

The company will provide a web portal called Rabbit Hole where users can connect their service accounts. Rabbit says credentials will not be stored on the device and that users can remove the connections whenever they want.

Screen, wheel and rotating camera

The R1 has a 2.88-inch touchscreen, a scroll wheel and a physical push-to-talk button. It also includes a camera that can rotate to face forward or backward, which Rabbit calls Rabbit Eye. The camera is intended for visual tasks, although voice is the primary mode of interaction. Primary source.

Inside is a MediaTek Helio P35 processor, 4 GB of RAM and 128 GB of storage. It connects to 4G mobile networks and Wi-Fi, so it is not designed as an accessory that has to remain tethered to a phone. Rabbit has announced that it will not charge a monthly fee to use the device. Primary source.

Price is also part of its strategy. At $199, the R1 costs considerably less than a midrange phone and sits well below mixed-reality headsets and other experimental AI hardware devices. Rabbit expects to begin shipping in March. Primary source.

The promise depends on real-world reliability

The idea of a device that handles tasks for us is appealing because it targets an everyday frustration: many digital tasks are repetitive and require users to switch between services, passwords and screens. If the R1 can understand a request and complete every step correctly, it could save time on simple actions.

But that ambition is also its toughest test. A conversation can tolerate an imperfect response; hailing a car to the wrong location, ordering the wrong product or making an incorrect payment cannot. For the action model to be useful, it will need to interpret requests accurately, adapt to changes in app interfaces and make clear to users what action it is about to take.

Digital services also frequently change their designs, terms and access mechanisms. The R1’s value will depend on Rabbit keeping that learning up to date and making the experience more convenient than taking out a phone and opening an app.

The launch comes as artificial intelligence begins moving beyond chat windows and into physical products. Rabbit is not proposing an assistant that holds better conversations, but one that does specific things. In the coming months, the R1 will have to show whether that difference is enough to justify carrying a second device in your pocket.

Acting requires more than understanding a sentence

A request such as booking travel contains intent, data, constraints and financial consequences. The system must decompose it, show a plan, request missing fields and confirm before an irreversible step. A model navigating an interface can press the wrong button as fluently as a text model fabricates a fact.

Build tests with controlled accounts and reversible transactions. Measure whether the device selects the right service, preserves quantities, detects interface changes, handles an error and explains its action before execution. Success is not reaching the end of a demo; it is satisfying intent without extra actions.

Credentials are part of the product

Operating applications involves sessions, permissions and personal data. Users need to know where they are stored, who can access them and how to revoke them. Giving one portal a password may widen the harm from failure. Safe architecture uses limited authorisation, action logs and service separation.

An extra device also adds battery, connectivity, updates and support. Compare it with a phone by performing the same task and counting total time, corrections and cost. Purchase price is only one part; continuity depends on Rabbit maintaining integrations as applications change.

The camera should be evaluated through provenance and consent. Pointing it at an object for identification differs from capturing a person or document. Visible indicators, minimum processing and verifiable deletion keep a visual feature from becoming accidental surveillance.

The transferable skill is to audit any “agent that acts” through intent, plan, permission, confirmation, outcome and reversal. That path shows whether R1 reduces work or merely hides a fragile chain behind a spoken command.

Exceptions reveal agent quality

Demonstrations usually follow the happy path: available service, stable interface and complete request. A real test changes a price, removes an option, requires extra authentication or interrupts the network. The device should stop, explain state and offer alternatives without repeating a purchase. Exception handling is central to acting.

Learned preference must also be separated from intent. If the system remembers habits, users should be able to see, correct and delete them. A historical choice does not permanently authorise a purchase or recipient. Useful personalisation relies on controllable memory rather than invisible inference.

Voice adds ambiguity. Test noise, similar names and quantities. Before a financial action, the system repeats critical fields on screen. That redundancy may seem less futuristic, but it prevents incorrect transcription becoming a correct transaction for the wrong instruction.

The final report includes completed tasks, human steps, erroneous actions and reversibility. Pleasant conversation is scored separately. An agent earns trust by making fewer promises and leaving a better trace.

An update policy defines how long the device receives support and what happens if Rabbit stops maintaining an integration. Users should be able to export data and unlink accounts. A physical agent’s value depends on a continuing service; that dependence belongs to the purchase from day one.

The final comparison includes not automating. Some rare or delicate tasks cost less when users keep direct control than when they maintain an agent and review exceptions.