Codex Composer Predictions: Setup, Cost & Troubleshooting

Quick answer: OpenAI launched Composer predictions in Codex in beta on October 9, 2026. After Codex finishes a response, it may suggest a complete next message in the composer. Press Tab to place that suggestion in the message box, review or edit it, and then send it yourself. It does not send automatically.[2]

What are Composer predictions in Codex?

Composer predictions are optional, context-aware suggestions for the next instruction you might give inside an existing Codex conversation. The feature uses the conversation in the current thread. It can use memory or connected-app information only when that context is already present in the thread; it does not independently search memories or connected apps to create a suggestion.[1]

This is different from autocomplete. During the beta, Codex suggests a complete next message after a response rather than completing a sentence word by word while you type. It is also different from the prompts shown on a new-conversation screen: those help start a task, while Composer predictions continue an existing one.[1]

Availability: who can use it?

Requirement	Current beta status
Account	Personal ChatGPT Pro
Age	18 or older
Regions	All supported regions
Client	Latest Codex desktop app
Thread types	Local and SSH threads
Supported models	GPT-6 Astra and GPT-6.1 Sol
Chat on web/desktop	Not available there
Other personal plans	Not included in the current beta announcement

OpenAI notes that predictions may appear in local Work threads and conversations, but its current FAQ describes the beta as available only on personal ChatGPT Pro plans in the Codex desktop app. Treat any Work appearance as rollout-dependent rather than guaranteed.[1]

How to turn on and use Composer predictions

1. Update the Codex desktop app. An older build may not include the beta.
2. Sign in with an eligible personal ChatGPT Pro account.
3. Open a local or SSH thread. Cloud-only and ordinary Chat surfaces are not listed as supported for this beta.
4. Select GPT-6 Astra or GPT-6.1 Sol.
5. Open Settings → General → Composer and make sure Show predictions is on. OpenAI says it is enabled by default for eligible users.
6. Send a normal Codex instruction and wait for Codex to finish responding.
7. Look at the message box. A suggested next message may appear after several seconds; long threads can take longer.
8. Press Tab to accept the text into the composer.
9. Review or edit it before sending. Accepting a prediction does not submit the message.

There is no requirement to wait for a prediction. Start typing whenever you know what to ask next; Codex lets you ignore a suggestion and continue normally.[1]

How to turn predictions off

1. Open Settings in the Codex desktop app.
2. Select General.
3. Open the Composer section.
4. Turn off Show predictions.

Use the same switch to enable them again. The setting controls next-message predictions, not the starter suggestions displayed for new conversations.[1]

Do Composer predictions cost extra?

OpenAI says generating a prediction does not consume Codex credits or count toward Codex usage limits during the beta. There is no additional beta charge beyond the required eligible plan. However, when you send a message—including text accepted from a prediction—the resulting Codex work follows normal usage limits and billing.[1][2]

Action	Beta usage treatment
Codex generates a suggested message	Does not count toward limits or consume credits
You press Tab to place it in the composer	Does not send it
You send the accepted or edited message	Normal Codex limits and billing apply

Why predictions may not appear

Run these checks in order:

1. Confirm the plan: use a personal ChatGPT Pro account, not an unsupported personal tier.
2. Confirm the client: use the current Codex desktop app, not ordinary Chat on web or desktop.
3. Confirm the thread: open a local or SSH thread.
4. Confirm the model: select GPT-6 Astra or GPT-6.1 Sol.
5. Confirm the setting: make sure Show predictions is enabled.
6. Let the response finish: a suggestion can appear only after Codex completes its response.
7. Wait briefly: OpenAI says suggestions can take several seconds and may take longer in long conversations.
8. Try another natural turn: the composer intentionally skips some responses, so one missing prediction does not prove the feature is broken.

If all requirements are met and the feature repeatedly fails, contact OpenAI Support. OpenAI recommends including the suggestion or expected behavior, the model, and the desktop app version. Remove secrets and personal information from screenshots before attaching them.[1]

Privacy and safe-use checklist

• Read before sending. A prediction may not match your intent.
• Check scope words. Watch for suggestions that expand “inspect” into “change,” “test” into “deploy,” or “draft” into “send.”
• Keep secrets out of prompts. Do not accept a suggestion that asks you to paste tokens, credentials or private customer data.
• Verify destructive actions. Treat delete, reset, force-push, production-deploy and permission-change suggestions as high risk.
• Use repository safeguards. Keep version control, review diffs and run tests before accepting code changes.
• Turn the feature off if it distracts you. It is optional and does not continue a task autonomously.

Messages you actually send follow Codex retention and deletion policies. OpenAI also points users to ChatGPT data controls and its model-improvement policy for details. The prediction itself is not an automatic instruction to Codex; human review remains the approval step.[1]

Useful test workflow

Use a disposable repository or a clean test branch for your first session:

1. Ask Codex to inspect a small project and identify one harmless cleanup.
2. Let Codex respond, then note whether a next-message prediction appears.
3. Press Tab, but do not send immediately.
4. Check whether the proposed scope matches the task and whether it requests edits, commands or network access.
5. Edit the message to require tests and a diff review.
6. Send it, review the resulting changes, and verify the actual test output.

Implement only the agreed cleanup on this test branch.
Do not change dependencies, access the network, delete files or deploy anything.
Run the relevant tests and show the diff and actual test output for review.

This prompt is an editorial safety template, not an OpenAI requirement.

Frequently asked questions

How do I accept a Codex Composer prediction?

Press Tab to move the suggestion into the message box. Review or edit it, then send it yourself. Tab does not automatically submit the message.[1]

Why does a prediction not appear after every response?

OpenAI says the composer may skip a prediction. A missing suggestion after one response is not evidence of a fault; continue typing normally or test another eligible turn.[1]

Can Composer predictions continue a task automatically?

No. They do not automatically accept or send follow-up messages. You remain responsible for reviewing and sending each instruction.[1]

Do they work in existing conversations?

Yes, if the conversation meets the account, client, thread and model requirements. OpenAI notes that predictions may take longer in long conversations.[1]

Do they require extra keyboard permissions on macOS?

No. OpenAI says Composer predictions do not require additional keyboard permissions. If macOS requests one, check which app feature is asking for it.[1]

Can a prediction finish text I already started typing?

Not in the current beta. It suggests a complete next message rather than updating word by word. You can accept the complete suggestion and edit it, or ignore it and keep writing.[1]

Last checked: October 10, 2026 at 11:15 UTC. This is a beta feature, so eligibility, supported models and usage treatment can change.

Sources

1. OpenAI Help Center: Composer predictions in Codex
2. OpenAI Help Center: ChatGPT release notes