How to Use Codex Voice: Setup, Limits, and Fixes

Codex Voice works only when you start a new, empty task in Voice inside the ChatGPT desktop app. If you open an existing text task, the microphone is voice dictation, not a live conversation. Use the desktop app on macOS or Windows, confirm that your plan and workspace have Voice enabled, grant microphone access, and measure a short baseline session before delegating real work.

This guide shows the correct setup, explains what consumes Voice and Codex usage, and gives a safe diagnostic path for a missing Voice control, microphone failures, slow transcripts, 403 errors, and sudden disconnects.

Prerequisites

You need:

• the current ChatGPT desktop app on macOS or Windows;

• a ChatGPT Plus, Pro, Business, Edu, or Enterprise plan with the feature available to your account;

• Voice and Codex enabled by your workspace administrator, if your organization manages those settings;

• microphone permission for ChatGPT;

• a new, empty chat or task.

Availability can still depend on rollout status and workspace policy. Standalone Voice in Work and Codex is not available on the web or directly on mobile. OpenAI documents paired iOS Remote as the supported mobile path.

Step 1: Start Voice in the Correct Place

1. Open the ChatGPT desktop app.

2. Choose Codex.

3. Create a new, empty task.

4. Select Start new voice chat before sending any text.

5. Allow microphone access and choose a voice when prompted.

6. Say: "Tell me which project and task you can see, but do not edit or run anything."

Expected result: Codex answers aloud, a live transcript appears, and it identifies the active context without starting work.

If you already typed into the task, start another empty task. A text-first task can offer dictation, but it does not become a live Voice task. You can resume a task that originally started in Voice by opening it and selecting Start voice chat.

Step 2: Measure a Short Usage Baseline

Voice time and delegated task work are not the same meter.

• Connected Voice time is metered separately where flexible pricing applies.

• Work that Voice starts in Codex draws from the shared Codex agentic allowance.

• Ordinary Chat Voice uses separate caps and does not consume Codex usage.

• For credit-based Business and Enterprise workspaces, OpenAI currently documents Voice in Work and Codex at approximately six credits per minute. Do not apply that number to a personal plan.

Before a long session:

1. Open the Codex usage page and record the visible balance or limit state.

2. Run a two-minute Voice conversation without starting a task.

3. End Voice instead of leaving it connected in the background.

4. Refresh the usage page and note any change.

5. Start one small read-only task, then compare usage again.

Expected result: you have an account-specific baseline that separates connected conversation time from delegated agent work. If your plan exposes no exact usage number, record the banner state rather than estimating.

Step 3: Give Voice a Bounded Brief

Voice is most useful for intent, decisions, and coordination. Exact identifiers, commands, and acceptance criteria are easier to verify in the transcript.

Use this structure:

Goal: inspect the current test failures.
Scope: read only; do not edit files or install packages.
Evidence: run the repository's narrow test command and quote the failing test names.
Stop condition: return before proposing a fix

Expected result: Voice creates or steers a bounded task, reports progress, and stops with evidence instead of silently expanding the request.

Correct a mistranscribed filename, flag, or environment immediately. Ask Codex to repeat the exact token before it acts. For risky work, keep approvals and irreversible decisions in text so you can review the precise command.

Step 4: End the Session Deliberately

Muting is useful when you need silence, but OpenAI does not document muting as a guaranteed pause in connected-time metering. Use End when the conversation is over.

Before ending, ask:

Summarize every task you started, its current state, and anything waiting for my approval.
Do not start new work

Expected result: the transcript contains a compact handoff and no task is left running only because Voice moved on.

Fix a Missing Codex Voice Control

Check these in order:

1. Confirm that you are in the ChatGPT desktop app, not the web app, CLI, or IDE extension.

2. Create a new, empty Codex task. A text-first task shows dictation instead of live Voice.

3. Confirm that the signed-in account has an eligible plan and that the rollout reached it.

4. In an Enterprise or Edu workspace, ask an admin to confirm that Advanced voice capabilities and any required early-access setting are enabled.

5. Update the desktop app, fully quit it, reopen it, and sign in with the same account and authentication method.

6. Compare a new Chat Voice conversation with a new Codex Voice task. Record which control is absent.

Expected result: either Voice appears, or you have isolated the failure to account eligibility, workspace policy, rollout state, or the Codex surface. Reinstalling repeatedly does not prove which gate failed.

Fix Microphone Permission Errors

On macOS:

1. Open System Settings > Privacy & Security > Microphone.

2. Enable ChatGPT.

3. If you plan to share computer context, also review Screen & System Audio Recording and Accessibility.

4. Fully quit and reopen ChatGPT after changing permissions.

On Windows:

1. Open Settings > Privacy & security > Microphone.

2. Turn on microphone access and access for desktop apps.

3. Confirm that the intended input device is the system default.

4. Fully quit and reopen ChatGPT.

Start a new empty Voice task and say a short sentence. Expected result: the input indicator responds and the transcript matches the selected language. If the wrong language is detected, set the ChatGPT language explicitly instead of repeating the same phrase louder.

Diagnose Disconnects, 403 Errors, and Slow Transcripts

Recent public issue reports include Connection reset by peer (os error 54), a WebSocket sideband request blocked with a 403 challenge, and transcription slowing after several minutes. Treat those as different failure classes.

Use this isolation sequence:

1. Check the OpenAI status page for a current incident.

2. Record the desktop app version, operating system, exact UTC time, and visible error.

3. Retry once in a new empty Voice task.

4. Test another trusted network without a corporate proxy, VPN, TLS inspection, or content filter.

5. If the alternate network works, give your network administrator the timestamp and error. Do not disable managed security controls permanently.

6. If every network fails, attach the minimal reproduction and version details to an existing matching issue or support request.

Expected result: you can distinguish a service incident, account or rollout gate, local permission problem, and network path failure. Do not delete local task history or application data as a first diagnostic step.

For a conversation that slows over time, end it and reproduce with a new Voice task while noting the time-to-degradation. A consistent threshold is more actionable than "Voice is slow."

Use Voice With External Tools Safely

Codex Voice follows the same permissions as the tasks it directs. Voice does not make an external integration safer, grant missing access, or replace schema review.

For a reproducible Aident setup, paste this exact instruction into a text task first:

Follow https://aident.ai/SETUP.md

Then start a Voice task with a bounded test:

Confirm my Aident Loadout account and Vault status.
Find one connected read-only Action relevant to this project.
Inspect its current input schema, run it once with a harmless input, and report
the result. Do not perform writes, edit the repository, or ask me to paste a
provider credential into the conversation

Expected result: one current Action schema is inspected, one harmless result is returned, git status --short remains empty, and no provider secret appears in the transcript.

This is the same practical pattern behind Aident's Ollama networking guide: use the exact problem language, answer early, run a reproducible check, and explain what each failure means. For usage investigation outside Voice, use the Codex usage spike diagnostic.

Use Aident Loadout to connect one read-only Action. Measure success as one schema inspected, one harmless result verified, and no repository diff or credential copied into chat.

Sources

• ChatGPT Voice, OpenAI, accessed July 27, 2026

• ChatGPT Work and Codex, OpenAI, accessed July 27, 2026

• Using Codex with your ChatGPT plan, OpenAI, accessed July 27, 2026

• ChatGPT Voice FAQ, OpenAI, accessed July 27, 2026

• Network recommendations for ChatGPT errors, OpenAI, accessed July 27, 2026

• Codex Voice disconnects with connection reset by peer, OpenAI Codex issue tracker, July 26, 2026

• Codex Voice sideband receives a 403 challenge, OpenAI Codex issue tracker, July 26, 2026

• Codex Voice transcription slows after several minutes, OpenAI Codex issue tracker, July 26, 2026

Refresh this guide when OpenAI changes Voice availability, task-start behavior, usage pools, workspace controls, supported platforms, or desktop permission requirements.