Rabbit OS3 went generally available on 22 September 2026, and it is the clearest reversal any hardware-first AI company has made so far. The product that was supposed to replace your phone is now a piece of software that runs on the computers you already own, reaches you through the messaging apps you already use, and charges you nothing per month.
The company that shipped the r1 — the orange £160 gadget that drew more mockery than any consumer device of its year — has spent two and a half years building the thing the r1 was originally pitched as. Rabbit OS3 is that thing, minus the hardware requirement.
Founder and CEO Jesse Lyu framed it as the end of a long build: “We’ve spent two and a half years building toward a system you don’t operate, but instead just tell it the outcome.” The second quote in the release is the actual product claim. Rabbit OS3 “allows users to stop worrying about whether a task is technically possible, and start collaborating on whatever is on their mind.”
This article covers what Rabbit OS3 actually is, how the bring-your-own-key pricing works, what the five-device model means in practice, how it compares with the agent products it now sits beside, and which of the r1’s original promises this release finally keeps.
Table of contents
- What Rabbit OS3 Actually Is
- The Pricing Model Is the Real News
- What the Five-Device Model Means in Practice
- How Rabbit Got Here: r1, Intern, OS2
- What Rabbit OS3 Can Actually Do Today
- Setting Up Rabbit OS3: What the First Hour Looks Like
- How Rabbit OS3 Compares With the Rest of the Field
- What This Means If You Are Evaluating Agent Software
- The Sceptical Case Against Rabbit OS3
- The Cyberdeck and What Rabbit OS3 Is Really For
- Frequently Asked Questions About Rabbit OS3
- Rabbit OS3: The Verdict
- References
What Rabbit OS3 Actually Is
Rabbit calls it an agentic operating system, which is marketing language for something more specific: a cloud service that drives local machines through an agent you install on each one.
The architecture in one paragraph
Rabbit OS3 runs in the cloud. On each machine you want it to control, you install a local rabbit agent. The cloud service holds the conversation, the memory and the reasoning; the local agent executes on the device, reads local files without uploading them, and reports back. One account can connect up to five devices, and the service decides which device — or which combination of devices — should run a given task.
The single-thread interface
The design decision that distinguishes Rabbit OS3 from most agent products is that there are no sessions. Everything happens in one continuous conversation with persistent memory, across every connected device and every access channel. You can start a task in a browser on a desktop and continue it over SMS from a phone without re-establishing context.
DLAM and direct computer control
The company’s Large Action Model lineage survives in what it now calls DLAM, which combined with the local agents gives the system direct computer control. Rabbit says the system can write and debug its own workflows, which is the capability that separates an agent from a scripted automation tool.
| Layer | Where it runs | What it holds |
|---|---|---|
| Chat interface | Browser, Telegram, iMessage/RCS/SMS, r1 | One continuous thread |
| Orchestration | Rabbit cloud | Memory, routing, model calls |
| Local agent | Each connected device | Execution, local file access |
| Model provider | Your API key or local model | Reasoning, code generation |
The Pricing Model Is the Real News
Rabbit OS3 has no monthly subscription. You supply the model access yourself — an API key from a frontier lab, a cloud router platform, or a locally hosted open-weight model — and Rabbit charges you nothing for the orchestration layer.
Why bring-your-own-key matters here
The release wording is deliberately broad: keys from “the latest models of frontier AI labs, cloud router platforms or locally hosted models”. Rabbit OS3 explicitly supports swapping providers mid-conversation without losing context, which turns model choice into a runtime setting rather than a platform commitment.
What it costs you instead
Nothing is free; the cost moves. Instead of a flat monthly fee you pay per token to whichever provider you point it at, which means a heavy agentic workload can cost more than a subscription and a light one considerably less. The table below sets out how that compares with the prevailing pricing shapes.
| Model | Who pays the inference bill | Cost predictability | Provider lock-in |
|---|---|---|---|
| Flat monthly subscription | Vendor absorbs it | High | High |
| Credit or task packs | You, in advance | Medium | High |
| Rabbit OS3 bring-your-own-key | You, per token | Low | None |
| Local open-weight model | You, in hardware and power | High after purchase | None |
The commercial logic
A vendor that absorbs inference costs has to cap usage, and capped agents are bad agents — they stop halfway through long tasks. By refusing to sit in the inference supply chain at all, Rabbit removes its own incentive to throttle. It also removes its most obvious revenue line, which is the open question hanging over the launch.
What the Five-Device Model Means in Practice
The device limit is the most concrete number in the announcement, and it shapes what Rabbit OS3 is actually for.
Which devices count
Windows, Mac and Linux PCs all count, as do cloud virtual machines, dedicated AI machines, the r1 itself, and the cyberdeck Rabbit says will ship with OS3 as its default. Five is enough for a personal fleet — a work laptop, a home desktop, a cloud box, a phone-adjacent device and one spare — and not enough for a team deployment.
Distributed execution
Because the orchestration layer knows about every connected device, it can split a task. A long-running job can sit on the cloud VM while a file-heavy step runs on the machine that actually holds the files, with results reconciled into the same conversation. That is the part conventional desktop agents cannot do at all.
The privacy position
Rabbit’s stated boundary is that conversations and memory live on its servers while local files stay on your machine — the local agent processes them in place rather than copying them up. Actions are user-initiated, sensitive operations require confirmation, and permissions can be revoked. It is a reasonable position that still requires trusting the cloud with the thread itself.
That last bar is why Rabbit OS3 exists. Five thousand daily actives against roughly one hundred thousand purchasers is a five per cent daily engagement rate, and no hardware roadmap survives that number.
How Rabbit Got Here: r1, Intern, OS2
Rabbit OS3 is the fourth attempt at the same idea, and the history explains the design.
The r1 and the Large Action Model
The r1 launched at CES in January 2024 on the promise of a Large Action Model that would operate apps on your behalf. It sold roughly 130,000 units. Reviews were brutal: Marques Brownlee called it “barely reviewable”, The Verge said it had “launched before it was ready”, and TechCrunch described it as “a fun peep at a possible future” while advising readers not to buy one.
The Android question
Android Authority demonstrated that the r1’s software ran on an ordinary Pixel handset from a leaked package file, and The Verge confirmed it. Lyu acknowledged the Android base while rejecting the characterisation that the r1 was merely an app. The episode fixed a reputation the company has been arguing with ever since.
Intern and rabbitOS 2
In June 2025 Rabbit shipped intern, its first software product — a general agent that produced websites, presentations and small tools, with the first three tasks free. rabbitOS 2 followed in September 2025 with a card-based interface and a “creations” feature for generating tools from natural language prompts. Both were steps toward decoupling the agent from the device.
The hardware exit
Rabbit has stopped manufacturing the r1 and ruled out an r2. The next hardware is a cyberdeck aimed at coding, described as months away rather than quarters. For a company whose original thesis was that hardware would replace app-based operating systems, shipping a cross-platform software agent and retiring the device is a complete inversion.
| Release | Date | Form | Core claim |
|---|---|---|---|
| r1 | January 2024 | Handheld device | Large Action Model operates your apps |
| intern | June 2025 | Web software | General agent produces finished artefacts |
| rabbitOS 2 | September 2025 | Device firmware | Card interface and generated tools |
| Rabbit OS3 | September 2026 | Cross-platform service | One thread drives five machines |
What Rabbit OS3 Can Actually Do Today
The announcement describes capabilities rather than benchmarks, so the honest framing is what the design permits rather than what has been measured.
Research and document work
The intern lineage covered research, presentations, reports and small interactive sites, and Rabbit OS3 inherits that. The difference is that output now lands on a real filesystem on a real machine rather than in a web workspace, which removes the export step that made the earlier product awkward for anything beyond a one-off artefact.
Local file operations
This is the capability that browser agents cannot match. Because the local agent reads files in place, Rabbit OS3 can work across a folder of documents without any of them leaving the machine — renaming, extracting, summarising, reconciling. Most of that work is ordinary natural language processing applied to text the model can see, and the interesting part is not the analysis but the fact that the analysis happens where the files already are.
Writing and debugging its own workflows
Autonomous coding is listed as a first-class capability, not a side effect. In an agent context that mostly means the system writes a script when a script is the right answer, runs it, reads the error, and fixes it — a loop that determines whether long tasks finish or stall.
Cross-device handoff
The distinctive task shape is one that touches two machines: pull data on the cloud box, join it against files that only exist on the laptop, produce output on either. That is awkward to script, tedious to do by hand, and it is the scenario Rabbit OS3 is built around.
Setting Up Rabbit OS3: What the First Hour Looks Like
The onboarding sequence is short, and each step carries a decision worth thinking about before you click through it.
Sign in and connect the first device
Access begins at the OS3 web portal on a desktop. The local agent installs on the machine you want controlled, and that machine becomes the first of your five. Start with a device that holds nothing sensitive; the permission model is easier to reason about once you have watched the agent work.
Choose a provider and supply a key
This is the consequential step. A frontier-lab key gives the best capability and the least predictable bill; a cloud router key gives fallback across providers; a locally hosted open-weight model gives a fixed cost and lower capability. Because Rabbit OS3 allows swapping providers without losing the thread, this decision is reversible, which is unusual and worth exploiting during evaluation.
Add skills deliberately
The universal skill mechanism accepts any skill URL pasted into the chat and configures itself from it. Treat pasted skills the way you would treat a browser extension: each one extends what the agent will do on your machine, and the convenience of one-line installation is exactly what makes an unreviewed skill a problem.
Set the confirmation boundary
Sensitive actions require confirmation and permissions can be revoked, so decide early which categories you want to approve by hand. Purchases, outbound messages and anything that writes outside a working directory are the sensible defaults for manual approval.
How Rabbit OS3 Compares With the Rest of the Field
Rabbit OS3 arrives into a category that did not exist when the r1 launched, which is the single biggest reason it has a chance.
Against desktop computer-use agents
The major labs now ship agents that operate a computer, but they are generally bound to one machine and one provider. Rabbit OS3’s multi-device orchestration and provider independence are genuine differentiators; its weakness is that it does not own the model, so capability tracks whatever key you supply.
Against browser-resident agents
Browser agents are easier to trust and much narrower. They cannot touch local files, drive desktop applications or coordinate work across machines. Rabbit OS3 trades that safety margin for reach, which is exactly the trade an intelligent automation programme has to make deliberately rather than by accident.
Against scripted automation
Traditional automation is deterministic and brittle; it does precisely what you wrote and fails when a page changes. Rabbit OS3 writes and debugs its own workflows, which makes it adaptive and correspondingly harder to audit. Neither property is strictly better, and serious deployments will want both tools.
The skills question
The universal skill mechanism — copy a skill URL, paste it into the chat, and the system configures itself — is the least proven part of the launch. Skill portability has been promised by several agent platforms and delivered by none at scale, largely because skills encode assumptions about the environment they were written in.
What This Means If You Are Evaluating Agent Software
Rabbit OS3 is worth a look for a specific set of readers and safe to ignore for most.
Who it fits
Individual technical users with several machines and an existing API key get the most out of it, because the cost model rewards people who already pay for inference and the device model rewards people with a fleet. Developers are the obvious first audience, which is what the cyberdeck positioning confirms.
Who it does not fit
Teams needing shared administration, audit trails, single sign-on and per-seat billing will find nothing here. Five devices per account is a personal-scale limit, and there is no stated enterprise tier. Organisations comparing agent platforms against a broader AI strategy should treat this launch as a signal about pricing direction rather than a procurement candidate.
The questions to ask before connecting a work machine
Ask what the local agent’s permission scope actually is, whether your employer’s policy permits a third-party cloud service to hold a persistent thread containing work context, and what happens to that thread if you revoke access. Those are answerable questions, and the answers are not yet public in detail.
The Sceptical Case Against Rabbit OS3
There are three reasons to hold judgement, and all three are about the company rather than the software.
No stated revenue model
A product with no subscription and no inference margin has to make money somewhere. The plausible answer is the cyberdeck, which makes Rabbit OS3 a funnel for hardware — the same bet that produced the r1’s engagement numbers, running in the opposite direction.
The demo-to-delivery gap
Rabbit’s history is of capable demonstrations that arrived incomplete. The r1’s Large Action Model was announced in January 2024 and the web-based agent reached devices in late September that year. Judging Rabbit OS3 on its launch materials would repeat the mistake reviewers made the first time.
Dependence on provider terms
Bring-your-own-key is only durable while providers permit third-party orchestrators to drive their models for computer-use tasks. Terms change, and a platform with no model of its own has no fallback beyond locally hosted open weights, which lag the frontier by a meaningful margin.
The Cyberdeck and What Rabbit OS3 Is Really For
The hardware Rabbit has not stopped making tells you how the company reads its own launch.
A coding machine, not a companion device
The cyberdeck is described as a device for coding, shipping with OS3 as its default environment. That is a very different customer from the r1’s: someone who already owns several machines, already holds an API key, and already writes software. It is also a customer who can evaluate an agent honestly, which the r1’s audience largely could not.
Why the free layer makes sense in that light
If the revenue is in the cyberdeck, Rabbit OS3 is a demand-generation layer and giving it away is rational. The risk is the one every hardware-funded software platform carries: if the device does not sell, the free layer becomes a cost centre, and cost centres get subscription tiers bolted onto them.
The install base problem it solves
Shipping to Windows, Mac and Linux means Rabbit OS3 can reach millions of machines on day one rather than the hundred thousand or so r1 units in the world. That is the single largest strategic difference between this launch and everything the company has done before, and it is worth more than any individual feature in the release.
Frequently Asked Questions About Rabbit OS3
Do I need an r1 to use Rabbit OS3?
No. Rabbit OS3 runs on Windows, Mac and Linux machines and cloud virtual machines, and is reached through a browser, Telegram or iMessage, RCS and SMS. The r1 is one supported access point among several, not a requirement.
Is Rabbit OS3 free?
There is no monthly fee for the service itself, but you must supply model access — an API key from a provider, a cloud router, or a local model — and you pay whatever that provider charges for the tokens your tasks consume.
How many computers can it control?
One account can connect up to five devices. The orchestration layer chooses which device or combination of devices runs each task.
Does Rabbit upload my files?
Rabbit says the local agent processes files on the device without copying them to its servers. Conversations and persistent memory are stored on Rabbit’s servers, so the thread itself is in the cloud even when the files are not.
Is Rabbit still making the r1?
No. Manufacturing has stopped and the company has ruled out a direct successor. The next hardware product is a cyberdeck aimed at coding, which will ship with OS3 as its default environment.
Rabbit OS3: The Verdict
Rabbit OS3 is a better product than the r1 was, built on a cost structure that is genuinely friendlier to users than the subscription default, and it does one thing — coordinated execution across several machines from a single thread — that almost nothing else does. The company has also finally stopped asking anyone to carry an extra object.
The doubts are commercial, not technical. A free orchestration layer funded by future hardware sales is a fragile arrangement, and Rabbit’s record is of shipping the interesting half first. If you already pay for model access and run more than one machine, Rabbit OS3 costs you an evening to evaluate and nothing per month thereafter. That is a reasonable offer, and it is more than the r1 ever managed. For the wider shift this sits inside, see our coverage of AI security as an engineering problem at every layer, which sets out what connecting an autonomous agent to a working machine actually exposes.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
