Assistants that message users first, rather than waiting for a prompt, moved from concept to product in October. Anthropic opened a research preview of Claude Duet on 21 October, an app that reads connected email and calendar accounts, sends a morning briefing and keeps working while the user is away. Block unveiled Dessa on 23 October, a voice model trained only on human speech that listens and talks at the same time.
The two assistants point in different directions. One of the assistants, Duet, is built for office work, with tasks handed to it by text or voice in a single ongoing thread. Dessa is one of the assistants, a research model rather than a public product, and its claim is about how a voice system learns to talk. Together, they show how far the assistant market has moved from a chatbot that answers questions to software that watches accounts, starts work and speaks in real time.
Duet runs on Claude Opus 5.5, Anthropic’s top model, released in late September. It is a separate app from the main Claude app, available on iOS, Android and claude.ai. Access is by waitlist, and only for paying Pro and Max subscribers in the United States for now. Team and Enterprise plans are not part of the preview.
Assistants send a morning briefing before work starts
Each morning, Duet sends a short summary of what most needs attention across connected apps such as email and calendar. The summary is called the Claude Digest. During the day, it checks those accounts every few hours and messages the user when something needs a decision. Anthropic says it checks at intervals rather than constantly, because constant checking would use up a subscriber’s usage limits much faster.
Users can also hand Duet work by text or voice, from scheduling meetings to building a sales forecast, a slide deck or a simple web app. It has its own virtual machine, a computer in the cloud with its own browser, so it keeps working when a phone or laptop is off. It can use more than 6,000 connectors and plugins in Anthropic’s marketplace.
Assistants must ask before they send, pay or delete
Approval is the central design choice in Duet. It reads, drafts and prepares work on its own, but before it sends a message, makes a payment, signs a document or deletes anything, it shows the user exactly what it is about to do and waits. Some actions are blocked even with approval, including trading securities and moving money between accounts.
Anthropic says Duet never directly sees passwords or sign-in tokens, and a new classifier, an automated filter that checks each message, is designed to refuse to share the most sensitive data, such as card numbers and government identification. Health or financial details go out only on the user’s explicit instruction. Purchases go through Stripe’s Link wallet, which asks the user to approve the spend and pays with a virtual card. The preview is included in Pro and Max plans, which start at $17 a month and $100 a month in one account, and at $20 a month for Pro in another, with background work counting toward existing usage limits.
Assistants that message first are useful only if the user trusts them to stop before doing something irreversible.
Assistants learn to talk by listening, Block says
Dessa takes a different route to a similar goal for assistants: conversation that feels natural. Block says Dessa was trained from scratch on nothing but human conversation, with no text and no pretrained model. “Audio in, audio out,” Block’s Dessa page says. The model listens and talks at the same time, a design called full duplex, so it can cut in, offer short acknowledgments such as “yeah” or “got it,” and react to laughter, sarcasm and changes in tone.
Block says Dessa learned language from hundreds of times fewer words than large language models consume, and that the whole model was trained on a single Nvidia GB200 NVL72 system. It says inference runs faster than real time, and that the model runs locally on a MacBook Pro with an M4 Max chip or better. Jack Dorsey, Block’s co-founder, said: “Talking with it feels like talking with a person, full of emotion, nuance, pauses, and a natural cadence.”
The limits matter for both assistants. Duet can miss something urgent that arrives between its checks, and its background work counts against usage limits. Dessa has shown selected recordings, not a general service, early access goes first to safety researchers, and its training claim has not been backed by published data. The next evidence for both will come from wider use: whether users accept assistants that message first, and whether a voice model trained only on speech holds up outside company demonstrations.