ChatGPT Business Adds Voice Control for Work and Codex Agents
OpenAI is turning voice from a conversational feature into a control surface for starting, steering and checking delegated agent work.
OpenAI has added voice control for Work and Codex to ChatGPT Business, moving speech beyond question-and-answer conversation into the operation of delegated agents. Users can start tasks, ask what an agent is doing, check progress and coordinate several agents through one spoken conversation.
The release separates two experiences that sound similar but serve different purposes. Voice in Chat is the conversational mode powered by GPT-Live and is available in desktop Chat as well as supported web, iOS and Android experiences. Voice in Work and Codex is the agent-control layer.
Where agent voice works
Voice in Work and Codex is available in the ChatGPT desktop app on macOS and Windows, with paired access from an iPhone. OpenAI says the standalone agent-control experience is not available on the web or mobile.
That platform boundary reflects the job the feature is doing. A normal voice conversation ends when the exchange ends. Agent work can continue in the background, move across tasks and require interruption or redirection. Voice becomes a way to supervise that process without keeping the agent transcript at the centre of attention.
The practical appeal is easy to see. A user could begin a research task, ask for an update while working elsewhere, redirect an agent that has taken the wrong path and check whether another task is ready for review. The interface is conversational, but the underlying work remains delegated and asynchronous.
How usage is metered
ChatGPT Business workspaces include one hour of Voice in Chat. OpenAI lists additional conversational voice use at five credits per minute. Voice in Work and Codex uses approximately six credits per minute.
Those voice credits pay for the control conversation, not the delegated work itself. Tasks carried out through Work or Codex continue to draw from the workspace’s shared usage pool at standard rates. A long agent task can therefore consume ordinary task usage while the spoken supervision layer accrues voice usage separately.
That distinction will matter for administrators. Voice may make agents easier to direct, but it also creates a new metered surface whose value depends on how much hands-free coordination replaces screen time rather than merely duplicating it.
Why it matters
Voice assistants have traditionally been designed around short commands and immediate answers. OpenAI’s release points toward a different role: voice as an operating layer for work that continues after the command is spoken.
The useful question is no longer whether an assistant can understand a request. It is whether a person can reliably supervise several ongoing tasks without losing context, control or a clear view of cost. Voice lowers the friction of issuing instructions; the quality of progress reporting and interruption will determine whether it also lowers the friction of management.
Status
Confirmed. OpenAI documents availability, supported platforms and Business credit rates in its official release notes. The productivity implications remain an editorial assessment, not a measured outcome.
Sources
Update note: Last reviewed 2026-07-24. We will revise this post if availability or Business credit rates change.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.