Table of Contents
- The Two Products Inside Orbit
- The Briefing Assistant Is the Expected Half
- The Phone Agent Is the Story
- What a Phone-Operating Claude Actually Changes
- The Risks Anthropic Has to Solve
- How to Get Access and What to Expect
- FAQ
Anthropic announced Orbit at its Code with Claude conference in San Francisco on May 6. Most of the coverage treated it as one product: a proactive AI assistant, Anthropic’s answer to ChatGPT Pulse. That framing is half right and it buries the more important half.
Orbit is two products under one name. One is a proactive briefing assistant for Claude Cowork that auto-generates daily insights from your connected work tools. The other is a mobile agent that operates your actual phone — placing calls, sending messages, navigating apps, managing your calendar by tapping and typing the way a person does. The first is a competent entry into a category that already exists. The second is the one that matters.
The Two Products Inside Orbit
Both halves of Orbit shipped to research preview on May 6. No public release date has been announced; references in May 2026 builds suggest a staged rollout.
The briefing assistant lives inside Claude Cowork and Claude Code. It is opt-in, time-zone-aware, and connects to Gmail, Slack, GitHub, Calendar, Drive, and Figma. It generates personalized briefings with actionable insights on a schedule, surfacing what changed in your work tools while you were away.
The phone agent is a separate capability. It runs on iPhone and Android and takes real actions on the device — calls, messages, app navigation, calendar management — by operating the interface directly rather than calling APIs. It is the mobile equivalent of computer use, applied to the device most people actually run their lives on.
Anthropic packaged them together because they share a thesis: Claude should act without waiting to be asked. But they are different products with different risk profiles, different competitors, and very different implications. Conflating them does both a disservice.
The Briefing Assistant Is the Expected Half
The briefing assistant is good. It is also not surprising. OpenAI shipped ChatGPT Pulse in September 2025 with the same core idea — overnight briefings stitched from chats, memory, and connected Gmail and Calendar. Google’s Proactive Assistance is expected to land at I/O. The proactive-briefing category was always going to fill in across all three labs.
Where Orbit’s briefing half differentiates is the connector set. Pulse leans on email and calendar — the executive’s tool stack. Orbit explicitly includes GitHub and Figma. That is a deliberate targeting decision: Orbit’s briefing assistant is built for the people who build things — developers, designers, product managers — not for the people who only read email. A briefing that tells you what changed in your repos and your design files overnight is a different product than one that summarizes your inbox.
It is a real advantage. It is also a connector list, not a new capability. The briefing assistant competes on integration breadth, and integration breadth is the kind of advantage that erodes as competitors add connectors. Useful, defensible for a quarter or two, not the durable story.
The Phone Agent Is the Story
A Claude that operates your phone is a different category of thing entirely.
Every prior AI assistant on a phone has been a chat interface. You type or speak, it responds, you act on the response. Even the most capable voice assistants have been a layer you talk to. Orbit’s phone agent is a layer that acts. It sees the screen, decides what to tap, taps it, reads the result, and continues — the full perception-action loop, on the device, against the real apps you already use.
The reason this matters more than the briefing assistant: briefings tell you things. The phone agent does things. “Summarize what happened in Slack overnight” is useful. “Reschedule my 3pm, text the three people affected, and book the conference room for the new time” is a different order of useful, and it requires an agent that can actually drive the apps. Orbit’s phone agent is the first credible attempt at that from a frontier lab applied to the personal device.
It also closes a gap in Anthropic’s stack. Anthropic has had computer use for the desktop since late 2024. The phone has been the missing surface — and for most people, the phone is the primary surface. An agent strategy that stops at the laptop is an agent strategy with a hole in it. Orbit’s phone agent fills the hole.
What a Phone-Operating Claude Actually Changes
Three concrete shifts, if the phone agent works as described.
The app-by-app workflow collapses. The current phone experience is: open app, do thing, close app, open next app, do related thing. A phone agent that spans apps turns a five-app errand into a single instruction. The friction that phones added back — the tapping, the switching, the re-entering of context — is the friction the agent removes.
The assistant stops being a destination. You do not open Orbit to use Orbit. You hand it a task and go do something else. This is the same shift happening across Anthropic’s agent products — the chat surface becomes a dispatch surface. On the phone, that shift is more consequential because phone time is the most contested attention in the world.
Accessibility changes meaningfully. A phone agent that operates apps by perception is, incidentally, one of the most powerful accessibility tools ever shipped. Anyone who struggles with the fine motor control or visual parsing that modern app interfaces demand gets an agent that does the operating for them. Anthropic has not led with this framing. It is the quietest large implication of the launch.
The Risks Anthropic Has to Solve
A phone agent is also the highest-stakes consumer agent surface anyone has shipped, and the research-preview label is doing real work.
The phone holds your messages, your banking apps, your authentication codes, your photos. An agent with the ability to tap and type on that device has, by definition, the ability to send a message you did not authorize, approve a transaction you did not intend, or navigate somewhere it should not. Computer use on a desktop is risky. Computer use on the device that holds your two-factor authentication is riskier.
The honest read on the research-preview status: this is not caution theater. A phone agent that gets confirmation flows, permission boundaries, and reversibility wrong is a phone agent that causes real harm to real users. Anthropic shipping this to preview rather than general availability is the correct call, and anyone deploying it before the guardrails are proven is taking on risk they should price honestly.
How to Get Access and What to Expect
Orbit is in research preview as of May 6, 2026. The briefing assistant ties to Claude Cowork and Claude Code; the phone agent runs on iOS and Android. Anthropic has not published a general-availability date or final pricing, and references in May builds point to a staged rollout rather than a single launch.
Do this first, if you get preview access: start with the briefing assistant, not the phone agent. Connect your read-heavy tools — GitHub, Slack, Calendar — and let it generate a few days of briefings. It is low-risk, immediately useful, and it teaches you how Orbit reasons about your work before you hand it the ability to act. Then, and only then, try the phone agent on a single low-stakes, fully reversible task — “find the cheapest of these three things in my shopping apps and tell me which” — something where a wrong action costs nothing.
Skip the phone agent entirely, for now, on any device that holds banking, brokerage, or primary authentication apps you cannot afford an agent to touch by mistake. The capability is real and the preview guardrails are unproven. Wait for general availability and a published security model before pointing it at anything that moves money.
FAQ
What is the difference between Orbit’s two halves?
Orbit is a briefing assistant and a phone agent under one name. The briefing assistant lives in Claude Cowork and Claude Code and auto-generates daily insights from connected work tools. The phone agent runs on iOS and Android and takes real actions on the device — calls, messages, app navigation. They share a thesis (Claude should act proactively) but are different products with different risk profiles.
How is Orbit different from ChatGPT Pulse?
ChatGPT Pulse, shipped September 2025, is a proactive briefing assistant — it tells you things. Orbit’s briefing half does the same with a builder-focused connector set (GitHub, Figma, not just email and calendar). But Orbit also includes a phone agent that acts on your device, which Pulse does not have. The phone agent is the real differentiator.
Can the Orbit phone agent really operate any app?
Orbit’s phone agent operates apps by perception — seeing the screen, deciding what to tap, tapping it — rather than through APIs. In principle that means it can navigate any app a person can. In practice, research-preview reliability will vary by app, and Anthropic has not published a compatibility list. Treat broad app coverage as the goal, not the current guarantee.
Is Orbit safe to use with banking or authentication apps?
Not yet. Orbit is in research preview and the security model — confirmation flows, permission boundaries, reversibility — is unproven. An agent that can tap and type on your phone can, by definition, take actions on sensitive apps. Wait for general availability and a published security model before using the phone agent on anything that holds money or primary authentication.
When will Orbit be generally available?
Anthropic has not announced a general-availability date. Orbit launched in research preview on May 6, 2026, and references in May builds suggest a staged rollout rather than a single launch. Final pricing has also not been published.
Does Orbit work with Claude Cowork?
Yes. The briefing assistant half of Orbit is built into Claude Cowork and Claude Code. The phone agent is a separate mobile capability. If you already use Claude Cowork, the briefing assistant is the part of Orbit that will reach you first.