Rabbit asks OS3 to compare itself, a day after the agent launch

Why it matters

Rabbit is shifting its agent strategy from a dedicated AI device toward software that can operate across customers’ existing computers, while leaving model costs to users. The launch’s central test is whether cross-device actions are reliable and worth those costs.

Rabbit inc. asked its newly released OS3 agent to make a comparison table about itself, in a post on Rabbit’s X account on September 23rd. The prompt followed OS3’s general release the day before. The post asks for a table; it does not, by itself, establish how the system answered or whether the result is an independent assessment.

The timing puts the small demonstration next to a much larger repositioning. OS3 is designed to work across connected computers and models, while Rabbit’s earlier public identity centered on the r1, a handheld AI device. Rabbit’s launch announcement describes OS3 as a cloud-based operating system that uses a locally installed Rabbit agent to access computers and route tasks across as many as five connected devices. The company says the service became generally available on September 22nd after an invite-only beta.

Rabbit founder and CEO Jesse Lyu (@jessechenglyu) said the company had spent two and a half years building toward a system users direct by stating an outcome. Lyu previously founded Raven Tech, a conversational AI company acquired by Baidu, according to Rabbit’s earlier r1 announcement. That history helps explain the continuity in Rabbit’s pitch: it has repeatedly tried to move the user away from navigating individual apps and toward describing tasks in natural language. OS3 carries that thesis from a dedicated device into a software layer that can reach a user’s existing computer.

Rabbit’s release notes list the product’s main pieces: a continuous conversation with memory, model choice through users’ own API keys, computer control through Rabbit’s agent, and access through the web, messaging channels or r1. The company says users can connect Windows, Mac and Linux computers, cloud virtual machines and the r1. It also says users can install skills by pasting a skill URL into chat and switch models without losing memory, skills or connected devices.

That broader footprint changes the product’s competitive frame. The r1 asked customers to carry Rabbit’s hardware as a separate interface. OS3 instead asks them to connect software they already use and pay for model access through a provider they choose. Rabbit’s support page says OS3 is free and users pay their selected model provider according to that provider’s API pricing. That avoids a Rabbit subscription for the service, while leaving model usage costs with the user. The company has not published a usage-cost comparison in the materials cited here.

The arrangement gives Rabbit a way to make the r1 a physical access point to a broader agent product rather than the sole reason to use Rabbit software. Rabbit says OS3 can also be accessed from a browser, Telegram and phone messaging, and that its upcoming cyberdeck will ship with OS3 as its default operating system. Those are company statements about availability and roadmap, not evidence of adoption or sustained task performance.

The comparison-table request is an apt public test of the product’s own framing: can a conversational system produce a useful account of what it does and how it differs? The post supplies no benchmark, criteria or independently checked comparison. A table generated by OS3 would show that the system can produce a formatted answer; it would not establish that its descriptions of competitors or its own abilities are accurate.

Rabbit’s launch announcement describes direct control of desktop software, local files and web pages, as well as autonomous coding workflows. It also says conversations and memory are stored on Rabbit’s servers, while files remain on the user’s computer; relevant content is processed through Rabbit’s servers and sent to the user-selected model provider. Those details make permissions, data handling and reliability central to evaluating the product, beyond the polished comparison a model might write about itself.

Rabbit spent its first product cycle arguing that a separate device could simplify digital tasks. OS3 places the same ambition inside computers and messaging tools people already have. The company’s prompt gives the new system a quick, self-referential assignment. Whether OS3 can complete ordinary multi-step work reliably, and at a cost users will accept, remains a more consequential measure than its ability to describe itself.