OpenAI's ChatGPT Work Lets Voice Control Your Emails, Slack, and Calendar

OpenAI has expanded ChatGPT Voice with access to the connector ecosystem used by ChatGPT Work, support for the GPT-6 model family, and voice control for documents, presentations, spreadsheets, email, calendars, messaging, code repositories, and browser tasks. Spoken requests can now initiate multi-step work across connected services.

Developers and teams can expose an integration through ChatGPT Work and make its tools available in typed and spoken sessions. That shared runtime also raises the stakes for permissions, confirmations, error handling, and model selection because a voice request may trigger changes in an external system.

Voice can call connected tools

During an active conversation, ChatGPT Voice can invoke plugins to read a Gmail thread, find an open calendar slot, post to Slack, or perform another supported action without requiring the user to switch apps.

The feature uses the connector layer behind ChatGPT Work. Its plugins rely on the Model Context Protocol, or MCP, which provides a common format for models to discover tools, submit structured requests, and receive results from services such as Gmail, Google Calendar, Slack, and GitHub. Integrations available to Work can therefore be triggered through voice when the user’s plan and permissions allow it.

Three models split the workload

Voice sessions can use one of three GPT-6 models, allowing users to balance capability, speed, and cost against the demands of each task.

OpenAI says its internal factuality evaluation used de-identified, real-world conversations in which users had flagged errors. On that test, Sol produced roughly half as many mistakes as its predecessor and reached Astra-level reliability at a lower cost. Because the result comes from an internal evaluation, it remains a vendor-reported measure rather than an independent benchmark.

Lower error rates carry added weight when a model can draft messages, modify schedules, or invoke external tools. Teams still need validation and approval policies that match the consequences of each action.

Workflows move into conversation

ChatGPT Work now supports voice on the web and mobile devices, giving spoken requests access to the same workplace context used by typed prompts. Supported workflows include:

• Dictating a specification while ChatGPT drafts the document.
• Building a presentation from information in a Slack thread.
• Creating a spreadsheet from figures found in email.
• Reviewing a calendar for conflicts and meetings that require preparation.
• Drafting changes to accept, decline, or reschedule events.

These workflows can combine retrieval, reasoning, content generation, and tool calls within one session. A request to reschedule a meeting, for example, may require reading the calendar, identifying viable times, preparing an update, and waiting for approval before applying the change.

Approvals stay in the loop

The voice interface inherits ChatGPT Work’s approval controls. Actions that require authorization pause for review, and ChatGPT drafts outgoing email for confirmation by default. Requests to send a message or delete a thread therefore stop at the relevant confirmation gate before the external service is changed.

Plus and Pro users can also configure Work on the web, iOS, and Android to react to events such as:

• A new Gmail message arriving.
• A monitored Slack channel receiving a message.
• A GitHub pull request changing.

Example automations include summarizing an email when registration opens or drafting next steps after a client posts feedback in Slack. Slack monitoring requires adding @ChatGPT to each channel involved, while any protected follow-up action waits for user approval.

Plans determine access

The voice update is rolling out globally in the latest ChatGPT app, while access to plugins, ChatGPT Work, and specific GPT-6 models varies by subscription.

For agent builders, voice now acts as another client for the same model-selectable, tool-using runtime. Integrations should apply the same tool schemas, authorization boundaries, confirmation rules, and failure handling across spoken and typed requests.