Claude Duet: Research Preview, Pricing and How the Proactive Agent Works

Anthropic announced Claude Duet on October 21, 2026, opening a waitlisted research preview of a proactive AI assistant. It can prepare work and continue tasks while you are away, rather than waiting for every new prompt. Access starts with US Pro and Max subscribers.

Claude Duet availability and pricing

The product page lists Duet as included in Pro and Max during the preview, sharing those plans' usage limits. Team and Enterprise are not included yet. There are standalone iOS and Android apps, plus access through claude.ai on a computer. Joining the waitlist is not immediate access.

Duet runs on Opus 5.5 and brings across your Claude memory, skills and connected apps. Background checks consume usage, as do longer assignments. There is no separate unlimited pool for the assistant simply because it runs outside the main chat interface.

That distinction matters when deciding whether to try it. A brief that saves a few minutes each morning and a large research assignment compete for the same subscription allowance. The useful comparison is how much completed work a plan supports, rather than how many messages fit into a chat.

How the agent keeps working

In the launch announcement, Anthropic describes a cloud virtual machine with a browser. Your phone or laptop does not need to remain on. Duet checks connected accounts every few hours, prepares a morning brief and sends notifications when work finishes or needs review.

The practical change is that context can become a reason to start useful preparation. A customer meeting might justify assembling a brief before someone remembers to request it. The difficult product question is deciding which preparation is worth doing, and when an interruption is worth the user's attention.

For a first workflow, a meeting brief is easier to evaluate than a broad instruction to manage a business. You can check whether the right meeting was identified, whether the brief uses relevant information and whether it arrived early enough to help. Those are concrete outcomes to judge before extending the assistant's responsibilities.

What requires approval

Anthropic says Duet can read and draft independently but waits for approval before sending messages, paying, signing documents or deleting information. Securities trading and account-to-account money transfers remain off-limits even with approval. A dedicated classifier and safety prompt are intended to limit sensitive-data sharing; those are safeguards, not a guarantee that mistakes cannot happen.

Preparation and execution need different levels of trust. A draft can be inspected before it changes anything outside the conversation. A sent message creates a commitment or reveals information immediately. Keeping that distinction visible is especially important when one assistant reads across several business systems.

What the preview still cannot promise

Anthropic acknowledges that Duet can misunderstand a thread or judge urgency poorly. Its periodic checks can also miss something that arrives between them, and users cannot yet see its browser. The announced 1Password sign-in integration is coming later, not part of the current capability set.

That makes the preview a poor substitute for a time-critical alerting system. It is better evaluated as an additional layer of preparation: something that can surface unfinished work and reduce the effort of returning to it, while established operational alerts stay in place.

The larger shift is toward assistants that compete on judgment between requests. A model can produce an excellent answer and still make an unhelpful assistant if it spends the allowance on low-priority work or interrupts too often. Duet's real test will be whether the work it chooses to prepare earns the attention it asks for.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn AI launches and product thinking into clear stories that developers and founders read and share.

Book a 30-minute call.