Claude Duet vs Gemini Intelligence: Who Holds the Gate?

Two agents, two very different ideas about who is in charge. Claude Duet is Anthropic's standalone personal agent, built on Claude Opus 5.5 and opened as a research preview on a waitlist. Gemini Intelligence is Google's agentic layer built into Android, announced on 12 May 2026. Set them side by side on capability and you get nowhere, because neither company publishes an independent task-completion benchmark. Set them side by side on control and the comparison becomes concrete: who initiates, where the agent lives, what it can reach, what gate stands before an action, what is refused outright, and what either vendor actually measures. Those six labels are applied to both products below, and every place a vendor publishes nothing is marked not published rather than filled in with a guess.

Everything about Claude Duet here comes from Anthropic's own product page at claude.com/product/claude-duet, read on 24 October 2026. Everything about Gemini Intelligence comes from two Google posts published on 12 May 2026: the main announcement by Mindy Brooks, VP of Product Management, and its security companion by Dave Kleidermacher, VP of Platforms Security and Privacy. Vendor claims are attributed in the sentence where they appear. No figure below is estimated.

Where the agent lives

The first difference is structural. Claude Duet is an app you open: a standalone client for iOS and Android, plus the web at claude.ai, where the agent runs tasks on a virtual machine with its own browser. That last detail is worth pausing on, because it means your laptop can be off while Duet keeps working. It also means the agent begins at the edge of your machine and reaches inward only through what you connect to it.

Gemini Intelligence does not begin at the edge. Google's frame in the announcement is that Android is transitioning from an operating system into an intelligence system, and the assistant is part of the OS rather than an app sitting on top of it. Screen context is the clearest illustration: long-press the power button over a grocery list and ask Gemini to build a delivery cart, or photograph a travel brochure in a hotel lobby and ask it to find a similar tour. Neither action requires switching to a separate agent client, because the assistant is already inside the surface you are looking at.

Google's rollout language, as published on 12 May 2026, is that features will roll out in waves starting with the latest Samsung Galaxy and Google Pixel phones this summer, and will become available across your Android devices including your watch, car, glasses and laptops later this year. That is a plan, not an inventory. Nothing in either post establishes which devices carry which feature on 24 October 2026, so this comparison does not claim that any particular phone has it today.

Who initiates

Claude Duet initiates. It sends a morning brief, drafts next steps, and then goes quiet until something needs you. Anthropic publishes the throttle on that behaviour too: Duet checks connected accounts every few hours rather than continuously, because checking nonstop would use up the plan's limits much faster. The agent's proactivity is therefore bounded by an economics the vendor states openly, which is a different thing from a rate limit that users discover for themselves.

Gemini Intelligence, as described in the announcement, initiates nothing on its own for automation tasks. Google's sentence is explicit: Gemini only acts on your command and stops the moment the task is complete, and all that is left for you is the final confirmation. Progress is trackable live through notifications while it works. The user is the trigger, the task terminates at completion, and there is no described idle loop in which the assistant decides to start something.

The security post does name proactive features, citing Magic Cue when it describes ambient data safeguarded by Private Compute Core, Private AI Compute and protected KVM. So command-driven is the right description of the automation surface the posts specify, not a claim that every Gemini surface on Android is purely reactive. Google's own three principles for the programme are explicit user control, comprehensive data protection and operational transparency, and the first of those sets the tone for the rest.

What it can reach

Claude Duet's reach is published as a number. It works with more than 6,000 connectors and plugins, and your memory, skills and apps from the Claude app are present from the first message. The number describes breadth, not authority: it tells you how many doors exist, while the approval gate described further down is what governs which ones the agent can walk through without you.

Gemini Intelligence's reach is published as a boundary instead. From the security post: Gemini can only access the apps you allow it to work in, and not the rest of your device. Within that boundary the announcement describes multi-step tasks across apps, automation Google says it spent months fine-tuning on the Galaxy S26 and Pixel 10 for food and rideshare apps, and the screen-context examples above. It also lists the other surfaces at announcement: Gemini in Chrome from late June for research, summarising and comparing, with auto browse for appointment booking and parking; Autofill with Google powered by Gemini Personal Intelligence, strictly opt-in; and Gboard speech turned into polished text.

Google does not publish a scope list of which apps can be automated. Anthropic publishes a connector count. One side gives you a magnitude and the other gives you a permission rule with no enumerated scope, and neither substitutes for the other. A larger catalogue is not a larger authority, and a narrow permission rule is not evidence that the set it governs is small.

The gate before an action

Claude Duet's gate is enumerated, and it arrives before the action. Anthropic's page says that before Duet sends a message, makes a payment, signs a document or deletes anything, it shows the exact action and waits. Four verbs, four moments, one behaviour: the agent stops mid-task with a specific artefact in front of you, not a generic confirmation. Approvals, logins and safety checks run in your own cloud computer, which the page offers as the reason sensitive details stay private.

Gemini's gate is designed in rather than enumerated per action. The security post states that Gemini is designed to require user confirmation before making purchases on your behalf, and the announcement's control sentence ends with the final confirmation as the step that remains for the user. Around that sit the opt-in controls: granular enable and disable per feature, and app automation toggles that Google says will arrive later this year. Autofill with Google is described as strictly opt-in, and the connection to Gemini can be turned on or off in settings.

The practical difference is where the stop lands. Duet stops between a draft and its consequence, four times named. Gemini stops at the end of a task and before a purchase, with feature-level permission deciding whether the task runs at all. Both are gates. Only one of them tells an operator in advance which four actions will trigger it.

What is refused outright

A gate you can pass is not the same as a boundary you cannot cross, and this is the sharpest divergence in the two documents. Anthropic publishes both. Duet's approval gate covers actions you may approve; separately, trading securities and moving money between accounts are described as off-limits even with your approval, and credit card numbers and government identification are refused at the message layer. That second category never reaches the approval prompt. The answer is not a button that says no; there is no button at all.

Google publishes no equivalent list. The security post names its three principles and describes opt-in per feature, an app allow-list and the purchase confirmation, but it does not enumerate actions that stay refused even when the user consents. Not published is the honest reading. That does not mean nothing is refused; it means a buyer cannot audit where the boundary sits, in the way a reader of Anthropic's page can read four verbs and two off-limits categories and know.

What is measured

Neither vendor publishes an independent task-completion benchmark. That is the whole of the performance comparison, and it is why this article does not rank the two agents. What each company does publish is closer to accounting than to measurement. Duet shares the plan's usage limits with the rest of Claude, and its background checks count toward them; larger tasks such as building a financial model or a deck need more model turns and more background work. Cost is expressed in plan limits consumed, not in a per-task price.

Gemini Intelligence's per-task cost is not published, and neither is a benchmark. The nearest published measure is structural: the permission model and the rollout wave. Those tell you what the system is allowed to do and when it is scheduled to arrive. They do not tell you how often it completes a task correctly, and no number in either vendor's posts does.

What the two plans actually promise

Read the two documents as commitments and they diverge on the same axis as everything else. Anthropic's is a narrow release with named exclusions: a research preview, a waitlist, Pro and Max subscribers in the US, and Team and Enterprise plans not included yet. A price for Duet itself is not published, nor are device requirements or any availability date outside the US. The scope is small and every gap is stated as a gap.

Google's is a staged release across a platform. The wave sequence is the commitment: the latest Samsung Galaxy and Pixel phones first this summer, then watch, car, glasses and laptops later in the year, with per-feature opt-in and app-level automation toggles following. What the posts do not supply is a device-by-device list for today. The plan is legible; the shipped inventory as of 24 October 2026 is not, and neither post claims otherwise.

That asymmetry is the thing to carry away. One vendor tells you exactly which actions stop the agent and which are refused even if you insist, and tells you almost nothing about how well it performs. The other tells you how its permissions are structured and when the rollout is planned, and publishes no refusal list at all. If your question is which agent is more capable, neither company has given you the evidence. If your question is which boundary you can describe to a colleague before you hand over an account, only one of them has drawn it.

If you are building the agent loop yourself

Both products are consumer surfaces, and the gate you get is the gate the vendor chose. Teams that need a different one usually build the loop themselves, and at that point the model becomes a line item you can put a number against. OrcaRouter serves more than 200 models behind one OpenAI-compatible endpoint and one key, with 0% markup, the provider's list price passed through, and automatic failover between them. Claude Opus 5.5, the model Duet runs on, is $4.00 per million input tokens and $20.00 per million output tokens, with a 1M-token context and cache reads at $0.20 per million.

We host models, not apps. Nothing here changes what Duet or Gemini Intelligence will let you approve, refuse or automate; those boundaries belong to Anthropic and Google respectively. What a router changes is the layer underneath, where the approval logic is yours to write and the model that executes an approved action is a choice rather than a fixed dependency.