Anthropic Claude Duet and Block Dessa: Two New Directions in Personal AI
Anthropic introduces Claude Duet, a proactive AI agent for professionals and small businesses, while Block unveils Dessa, a full-duplex speech model designed for natural, real-time conversations.
Anthropic Claude Duet and Block Dessa: Two New Directions in Personal AI#
AI development is moving beyond conventional chat interfaces toward systems that can independently manage tasks and communicate through more natural conversations. Two October 2026 announcements illustrate this direction: Anthropic's Claude Duet, a proactive personal AI agent, and Block's Dessa, a speech-to-speech model designed for real-time conversational interaction.
Anthropic introduced Claude Duet on October 21, 2026, as a research preview for professionals and small-business owners. Two days later, Block unveiled Dessa, a voice model that the company says was trained entirely on human conversation audio rather than text.
Although the technologies address different problems, both explore how artificial intelligence can become more integrated into everyday work. Claude Duet focuses on taking action across business applications, while Dessa investigates how machines can communicate without relying on conventional speech-to-text pipelines.
Neither announcement represents a universally available new foundation model. Claude Duet remains a restricted research preview, and Block has not announced general public access to Dessa.
What Did Anthropic Announce With Claude Duet?#
Claude Duet is a proactive AI assistant developed by Anthropic to help professionals and small-business owners manage ongoing responsibilities.
Unlike a conventional chatbot that waits for a user to submit a prompt, Claude Duet is designed to identify tasks requiring attention, prepare information, and continue working in the background.
Anthropic describes it as an agent that can:
- Prepare morning summaries using connected applications.
- Identify important messages and calendar events.
- Draft email responses for user approval.
- Find unpaid invoices and prepare payment reminders.
- Prepare payroll information for review.
- Organize contracts and business documents.
- Generate reports, presentations, and dashboards.
The central idea is to reduce the amount of manual coordination required to manage everyday business operations.
For example, a business owner could ask Claude Duet to review outstanding invoices, identify customers who have not paid, and prepare reminder messages.
Instead of manually reviewing every record, the user receives prepared work that can be checked before taking action.
How Claude Duet Works#
Claude Duet is designed around a proactive agent workflow.
A conventional chatbot generally follows a request-response pattern:
- The user submits a prompt.
- The model generates a response.
- The user decides what to do next.
An agent-oriented application can operate differently.
It can periodically inspect connected information sources, identify relevant tasks, prepare actions, and notify the user when intervention is needed.
Claude Duet combines several capabilities:
- Application connectivity.
- Background task execution.
- Information prioritization.
- Document and content generation.
- Human approval before consequential actions.
Anthropic says Claude Duet checks with users before acting on their behalf.
This approval mechanism is particularly important for business operations involving communications, financial information, contracts, and other sensitive activities.
However, proactive behavior does not guarantee that an agent will correctly understand every business situation. Anthropic acknowledges that Duet can misinterpret email threads, misjudge urgency, or miss information between checks.
Source: Anthropic's Claude Duet announcement.
Claude Duet Availability and Access#
Claude Duet is available as a research preview through a waitlist.
Anthropic offers access to eligible Claude Pro and Max subscribers in the United States through its product preview. Availability is not equivalent to a general commercial launch.
The application is available through:
- A standalone iOS application.
- A standalone Android application.
- The Claude web experience.
During the research preview, Claude Duet shares the user's existing Claude plan usage limits. Background activity and more complex tasks can consume additional usage.
Anthropic has not presented Claude Duet as a separate universally available API product.
Businesses considering it should also examine application permissions, data access, and the consequences of allowing an AI agent to work with business information.
Source: Claude Duet product information.
Block Introduces Dessa: A Different Approach to Voice AI#
Block's announcement focuses on a different challenge: making spoken interactions with artificial intelligence more natural.
Called Dessa, the model is a full-duplex speech-to-speech system. Block says it was trained from scratch using human conversation audio, without text-based training or pretrained models.
The announcement was made on October 23, 2026, by Block's AI research team, with CEO Jack Dorsey also highlighting the development.
The model is designed to listen and speak simultaneously rather than waiting for a complete spoken request before generating an answer.
This matters because human conversations involve more than words.
People communicate through:
- Pauses and interruptions.
- Changes in speaking speed.
- Intonation and emphasis.
- Laughter and other vocal expressions.
- Conversational acknowledgements.
- Emotional cues in speech.
Traditional voice assistants frequently convert incoming speech into text, process that text using a language model, and then generate spoken output.
That pipeline can lose some information about how something was said.
Dessa attempts to approach the problem directly through audio-based conversation.
Source: Block's official Dessa announcement.
What Is a Full-Duplex Speech Model?#
A full-duplex speech model can process incoming audio and produce outgoing audio simultaneously.
This differs from a conventional turn-based voice assistant, where the system typically waits for the user to finish speaking before responding.
Consider two interaction patterns.
Traditional voice pipeline:
- A microphone captures the user's speech.
- Automatic speech recognition converts audio into text.
- A language model processes the transcription.
- A text-to-speech system generates spoken output.
Full-duplex speech interaction:
- The system receives incoming audio.
- It processes the conversational signal directly.
- It generates speech while continuing to listen.
- It can respond to interruptions and changes in conversational timing.
The second approach creates opportunities for more natural interactions.
It may allow an AI system to respond to pauses, acknowledge a speaker, and manage overlapping conversational activity more effectively.
However, full-duplex operation alone does not prove that a model understands human emotions or intentions accurately.
Those capabilities require separate evaluation.
What Makes Block Dessa Technically Different?#
The most interesting aspect of Dessa is Block's reported training approach.
According to the company, the model learned conversational behavior from human speech rather than relying on text-based training or an existing pretrained language model.
Block argues that conventional voice systems lose important information when speech is converted into text.
For example, the sentence "I'm fine" can communicate different meanings depending on the speaker's tone, timing, and delivery.
A text transcript may preserve the words but lose some of those acoustic signals.
Dessa attempts to retain that information by working directly with speech.
Training and Hardware Claims#
Block has made several technical claims about the model:
| Attribute | What Block reports |
|---|---|
| Model | Dessa |
| Developer | Block |
| Model type | Full-duplex speech-to-speech |
| Training approach | Trained from scratch on human conversation audio |
| Text-based training | Block says no text was used |
| Pretrained models | Block says none were used |
| Training hardware | One NVIDIA GB200 NVL72 system |
| Local inference | MacBook Pro with M4 Max or better |
| Inference speed | Block says faster than real time |
| Public benchmarks | Not published |
| Public model weights | Not announced |
| General API access | Not announced |
These specifications are based on Block's announcement and related reporting. The company has not published a detailed technical paper, model card, parameter count, or independent benchmark results.
The hardware and performance claims should therefore be treated as company-reported information rather than independently established results.
How Dessa Could Change Voice AI Applications#
If its demonstrated capabilities translate into reliable deployments, Dessa's architecture could be relevant to several applications.
1. Conversational AI Assistants#
Voice assistants could become less dependent on rigid question-and-answer interactions.
Instead of waiting for users to complete every sentence, a full-duplex system could support interruptions, acknowledgements, and more continuous conversations.
2. Customer Support Automation#
Businesses could investigate speech-native customer support systems that respond to customers in real time.
Such applications would still require reliable knowledge retrieval, transaction verification, escalation mechanisms, and identity checks.
Natural-sounding speech alone is not sufficient for dependable customer service.
3. Accessibility Applications#
More expressive voice interfaces could provide alternative interaction methods for users who prefer or require spoken communication.
However, accessibility benefits will depend on language coverage, speech recognition accuracy, latency, and the ability to handle different speaking patterns.
4. Local AI Applications#
Block says Dessa can run locally on a MacBook Pro equipped with an M4 Max chip or better.
If independently confirmed, local inference could reduce dependence on remote inference services and potentially improve privacy and response times.
It could also make speech-based applications more practical in environments with limited internet connectivity.
However, developers still need information about memory requirements, supported operating systems, model distribution, and deployment constraints.
AI Agents vs Speech-Native Models: Understanding the Difference#
Claude Duet and Dessa are not direct competitors.
They address different layers of the AI application stack.
| Area | Claude Duet | Block Dessa |
|---|---|---|
| Primary focus | Proactive task execution | Natural speech interaction |
| Developer | Anthropic | Block |
| Main technology | AI agent application | Full-duplex speech model |
| Interaction | Text and voice | Audio input and output |
| Core capability | Background work and task coordination | Real-time conversational speech |
| Target applications | Business productivity | Voice-based AI experiences |
| Availability | Restricted research preview | Limited research access |
| Public API | No separate general Duet API announced | Not announced |
| Open weights | Not announced | Not announced |
The distinction is important for developers.
Claude Duet is closer to a complete application that coordinates tasks and connected services.
Dessa is a speech model that could become one component of a larger conversational AI application.
A future voice-based business assistant could theoretically combine a speech-native interface with an agent that manages calendars, invoices, and customer communications.
That is an architectural possibility, not an integration announced by either company.
What Developers Should Consider Before Building With These Technologies#
Both announcements introduce interesting engineering possibilities, but their current access restrictions limit immediate production adoption.
Agent architecture and permissions#
Applications built around proactive agents need clear permission boundaries.
Developers should consider:
- Which applications an agent can access.
- Whether it can modify or delete business records.
- Which actions require human approval.
- How actions are logged and audited.
- How the system handles incorrect or incomplete information.
For financial and business applications, approval workflows should be part of the architecture rather than an optional interface feature.
Speech processing and latency#
Speech-native applications have different engineering requirements from conventional text-based chatbots.
Important evaluation areas include:
- End-to-end audio latency.
- Interruption handling.
- Conversational turn-taking.
- Speech quality and intelligibility.
- Language and accent coverage.
- Background noise robustness.
- Memory and hardware requirements.
A model that sounds natural in a short demonstration may still struggle with noisy environments, unfamiliar accents, or extended conversations.
Voice security and impersonation risks#
Highly realistic AI-generated speech creates additional security concerns.
Potential misuse includes impersonation, fraudulent calls, social engineering, and attempts to bypass voice-based identity checks.
Organizations should not treat a familiar-sounding voice as proof of identity.
Voice authentication systems need additional verification mechanisms, while AI-generated speech applications should consider disclosure, consent, and abuse prevention.
Block has invited a limited group of AI safety researchers to examine Dessa. This is particularly relevant because the same conversational qualities that make speech AI more natural could also make deception more convincing.
Sources:
What Comes Next for Personal AI and Voice AI?#
Claude Duet and Dessa illustrate two different directions in the development of AI systems.
Anthropic is exploring how an assistant can proactively manage information and complete tasks across connected applications.
Block is investigating whether a model can learn conversational behavior directly from human speech and interact without a conventional text-based pipeline.
For software developers, these developments point toward applications that combine reasoning, tool execution, and natural interfaces.
However, the practical value of both technologies will depend on factors beyond their demonstrations.
For Claude Duet, the important questions include task reliability, permission management, business integration, and predictable usage costs.
For Dessa, developers need independent evaluation, technical documentation, broader access, and evidence of its performance across real-world conversational conditions.
Neither announcement establishes that these challenges have been solved.
The next important milestone will be seeing how both technologies perform outside controlled demonstrations and limited research previews.