ChatGPT and Claude Announced Desktop Voice Within Minutes of Each Other: Driving Codex vs Thinking Aloud

July 24, 2026Updated: July 25, 2026Articles

Early on 24 July 2026, OpenAI and Anthropic announced updates to their voice features for the desktop at almost the same moment.

In Japan time, Anthropic’s official Claude account posted on X at 4:34 a.m. and OpenAI’s at 4:43 a.m. — about eight minutes apart. The coincidence of dates aside, the announcements marked voice becoming a way to drive work on a computer rather than a conversation feature on a phone.

What the two companies added is not the same thing, however. ChatGPT is heading toward starting and steering agents running in Work and Codex by voice. Claude is heading toward thinking problems through aloud with Opus and Sonnet, then executing the conclusions through connected services such as Gmail and Slack.

Same day, different center of gravity

Comparison	ChatGPT	Claude
Main aim	Starting tasks in Work and Codex, checking progress, coordinating several agents	Thinking through hard problems aloud, then executing via connected tools
On the desktop	ChatGPT desktop app on macOS and Windows	Claude desktop and web; also available on mobile
Conversation model	Real-time conversation via GPT-Live; you can interrupt mid-sentence	Turn-based: you speak, Claude thinks, then replies
Connection to your work	The files, apps, development environments, and agents that Work and Codex hold	Connectors such as Gmail, Slack, and Google Calendar
Models	ChatGPT Voice runs on the GPT-Live line	Opus and Sonnet selectable alongside Haiku
Japanese	Conversation in Japanese, as with ordinary Voice	Supported languages including Japanese now officially documented; must be selected manually

Judged on appearance — a microphone button appearing in a desktop app — they look alike. The difference becomes clearer if you read ChatGPT’s as a feature for supervising work already running, and Claude’s as a feature for organizing your thinking aloud and then having tools used where needed.

ChatGPT can drive Codex tasks by voice

What matters in OpenAI’s announcement is not only that ordinary chat gained a voice. Voice arrived in Work and Codex in the ChatGPT desktop app, so AI agents already working can be operated by voice.

The official Help Center describes these operations.

Start a new task
Prioritize among several tasks or agents
Interrupt work in progress and change direction
Ask why something is stalled, or how far it has got
Receive completion and blocked reports by voice and on screen

You can ask Codex to investigate a bug, then, while looking at another window, add by voice: “do not change the design yet, just identify the cause”, “stop before deploying once the tests pass”.

More significant than not having to type long instructions is being able to direct traffic conversationally while several jobs are in flight. ChatGPT Voice allows natural interruption, so you can stop and correct the moment an explanation diverges from your intent.

Ordinary Voice and Work/Codex Voice have different roles

ChatGPT’s Voice has been available on the web and on phones for some time. The old Voice on macOS was discontinued in January 2026, so this update is also its return to the new ChatGPT desktop app as a voice feature for work.

In ordinary Chat, voice handles questions, discussion, web search, and idea generation. Work advances longer jobs such as research and document preparation; Codex handles code changes, tests, and Git operations. Which of these you are in determines the tools and permissions Voice has available.

Operating Work and Codex by voice on their own requires the macOS or Windows desktop app. An iPhone can be used through Remote, paired with the desktop, but this is not a feature for starting Codex Voice independently from the web or mobile.

Voice connection time and the usage consumed by tasks Work and Codex run are counted separately. Voice allowance remaining does not help if you have hit Codex’s limit, and vice versa. Available features and limits vary by plan, region, workspace settings, and app version.

Claude moves from fast conversation on Haiku to thinking with its top models

Claude has had Voice for a while too. This is not the first appearance of voice conversation; the substance is that Opus and Sonnet, which think through hard problems, have joined a Voice that ran on Haiku for speed.

The model can be changed mid-conversation. Voice uses a fast variant of the selected model, and the model from your immediately preceding text chat is the initial selection. That makes it easy to continue a written discussion aloud and then return to text.

Claude’s official blog lists uses such as these.

Rehearse an important conversation and get feedback on your delivery
Explain a client proposal aloud and have the holes in the logic found
Generate product roadmap ideas and test them against competitive research
Voice several hypotheses and sort out which hold up

Unlike ChatGPT’s real-time conversation, Claude is turn-based: it thinks after you finish speaking, then replies. Rather than a design optimized for response speed, it is built as a sounding board that takes in half-formed thinking and deepens it by asking questions back.

Claude can execute through connectors once you have finished talking

Claude’s Voice does not stop at organizing your thinking. It can use connected tools to carry out what the conversation decided.

Move a Google Calendar meeting back by thirty minutes
Turn a conversation about a client proposal into a one-page Canva document
Summarize today’s email and draft replies to the important ones

Claude asks permission before using a connector. New connections are added under Settings → Connectors on mobile, desktop, and web. The more tools available through MCP and connectors, the more work can follow a spoken request.

This Voice update ships as a beta to all plans. Free gets Haiku, one connected tool, and all supported languages; paid plans widen the models and tools available. Voice conversations count toward your normal usage limits.

Claude supports Japanese, but you select the language first

Anthropic listed support for English, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Brazilian Portuguese, and two variants of Spanish.

The thing to note is that Claude does not detect the language you are speaking automatically. To switch from English to Japanese, either tell it by voice to change, or select Japanese in Voice’s language setting. The language you set in ordinary Claude does not carry over to Voice.

It works on a computer, but Anthropic says a phone is the easiest place to use it at present. It serves for long sounding-board sessions at a desk, while being out of the house or having your hands full remain the main use cases.

My own split: plan with Claude, talk to ChatGPT about work in flight

If you use both, the existing division — specifications with Claude, implementation with Codex — extends to voice unchanged.

Claude when your thinking has not settled

At the stage of “I cannot articulate what the problem is” or “I want to compare the options aloud”, Claude, where you can converse with Opus or Sonnet, is the fit. Asking it to hold off on answers and ask questions back makes it easier to use as a sounding board.

ChatGPT when you want to start and supervise work

When the goal is already set and you want Work or Codex doing the actual job, ChatGPT is the fit. “Start the research”, “have another agent do a comparison too”, “stop that change” — work in flight can be adjusted by voice.

The difference is not only which voice sounds more natural. You have to choose on what happens after the voice is received: which model thinks, which tools it uses, and how far it can execute.

Start by giving both the same ten-minute discussion

Update the ChatGPT and Claude desktop apps to the latest version
Pick one everyday question that involves no publishing or deleting
Give both the same background and the same finishing condition
Talk for ten minutes and compare how they ask questions and how they close
If tools are to be used, check what you are approving before execution

For something like “compare three ideas for the next article and decide which is closest to the reader”, you can compare not only speech recognition but whether it organizes without cutting you off, whether it asks the questions it needs to, and whether it ends with something actionable.

Check permissions for files and the screen, not just the microphone

Voice on a computer involves broader permissions than plain speech input. Using the state of your computer in ChatGPT may require screen recording and accessibility permissions alongside the microphone. Using connectors in Claude gives access to email, calendar, and files.

Convenience is no reason to grant every permission up front. The basics are the same as for text: do not share your screen with API keys or personal data displayed, do not read secrets aloud, and keep a confirmation step before sending, deleting, or publishing.

OpenAI notes that Voice transcription may not match exactly what was said. Do not settle important dates, amounts, proper nouns, or what is about to be executed by voice alone; confirm them in text on screen.

Voice shifts from an input method to a way of operating work

Voice mode has until now carried the impression of a feature for talking to an AI on a phone. With OpenAI and Anthropic announcing desktop updates within minutes of each other, its role is starting to change.

In ChatGPT, voice becomes a control panel driving several agents. In Claude, spoken thinking is organized with a top model and the conclusion handed to a connected service.

The keyboard is not going away. Long specifications, exact figures, and code diffs remain easier to check in text. Discussion while thinking, correcting course mid-task, and checking progress with your hands busy are more natural aloud. The question ahead is less voice or text than which parts of your work go faster spoken.

References

Announcement times were derived from the UTC timestamps in the X post IDs and converted to Japan time. Features, supported environments, and availability follow official announcements and help documentation as of July 24, 2026. Because rollout is staged, when a feature appears may differ by account.