Rabbit's Pivot to Software Admits Hardware Was Never the Point

Why Rabbit Finally Abandoned the R1

We noticed Rabbit quietly killed the hardware dream this week. Jesse Lyu announced OS3, a cross-platform agent that runs on iOS, Android, Windows, and Mac with no R1 required. The original pitch was a dedicated device that would replace your phone apps with a natural language interface. Two years and roughly 100,000 units later, the R1 sits in drawers while Rabbit builds what they should have built first: software that works on screens people already own.

What OS3 Actually Does That the R1 Couldn't

The technical shift matters more than the form factor. OS3 uses a Large Action Model that learns any app's interface by watching you use it once, then replicates those steps on command. The R1 could only do this through a cloud VM with latency that made it feel broken. Running locally on your phone or laptop cuts that round trip. We are seeing sub-second responses in demos versus the 10 to 15 second waits that killed the R1's credibility.

The Humane Comparison Nobody Is Making

Humane's Ai Pin died because it demanded a new hardware paradigm including a projector, magnetic battery, and cellular plan before proving the software worked. Rabbit at least shipped a $199 device that existed. But both companies bet on hardware as a moat. OS3 admits the moat was never the plastic; it is the action model. Humane never got to this pivot. They burned $230 million trying to make hardware the product. Rabbit spent maybe $50 million and lived to pivot.

Why Cross-Platform Changes the Competitive Math

Every major player including OpenAI, Anthropic, and Google is building agents locked to their ecosystems. Operator lives in ChatGPT. Computer Use lives in Claude. Project Mariner lives in Chrome. Rabbit's bet is that users want one agent that works across all of them. OS3 can theoretically book an Uber on your phone, then move a file on your laptop, then reply to a Slack on your desktop. That interoperability is the only differentiation left when every model is commoditized.

The Distribution Problem They Still Haven't Solved

Here is what the announcement glosses over. OS3 needs deep OS permissions to see and control apps. On iOS that means Screen Time hacks or enterprise certificates. On Android it means accessibility services that Google keeps restricting. On desktop it means admin rights. Rabbit has no platform leverage, no app store, no OS, no preload deals. They are asking users to sideload and grant scary permissions for an unproven agent. That is a harder sell than the R1 ever was.

What This Signals About the Next Quarter

The pivot timing is not accidental. Apple Intelligence rolls out broadly this fall. Google's Astra demos keep improving. OpenAI's Operator is in limited preview. Rabbit needed to ship something before the platform owners make third-party agents redundant. We are betting the next 90 days bring partnership talks, maybe with Nothing, maybe with a Chinese OEM, because Rabbit's only path to distribution is riding someone else's hardware. The software alone will not reach critical mass.

The Real Question: Can an Agent Be a Product Without a Platform

This is the existential question for every agent startup. Rabbit proved you can build impressive action models. They have not proven anyone will pay for one that lives inside someone else's OS. The R1 at least had a subscription narrative with $199 hardware then maybe a service fee. OS3's monetization is unclear. Freemium? Per-action pricing? Enterprise licenses? Until that is answered, this is a tech demo with a waiting list, not a business. We are watching the pricing page, not the demo videos.

What We Are Tracking Before Calling This a Win

Three things decide if OS3 matters. First: does the action model actually generalize to apps it has not seen, or does it need per-app training like the R1? Second: can they maintain compatibility when Apple and Google inevitably break the accessibility APIs they depend on? Third: will users tolerate an agent that sometimes works and sometimes hallucinates a $500 Uber order? The R1 failed on reliability. OS3 gets one chance to prove software-only does not mean the same bugs, just faster.