Last updated: 10 October 2026, 11:50pm BST
- Codex can now suggest a complete next message after it responds.
- The beta is for adults on personal ChatGPT Pro accounts in the latest Codex desktop app.
- It works in local and SSH threads using GPT-6 Astra or GPT-6.1 Sol.
- Pressing Tab accepts the text into the composer; it does not send it.
- Predictions are free of Codex limits and credits during the beta, but sent messages count normally.
OpenAI has added Composer predictions to Codex, a beta feature that suggests what a user might want to ask or approve next after the agent finishes a response.
The feature was announced on 9 October 2026 in OpenAI’s official release notes. It is a small interface change with a potentially large effect on long agent sessions, where the next instruction is often predictable.
How Codex Composer predictions work
After Codex replies, a suggested complete message may appear in the message box. Pressing Tab places that suggestion into the composer. The user can edit it, ignore it or send it.
The important distinction is that accepting is not sending. The beta does not automatically continue a task or authorise another action. OpenAI also says a prediction will not appear after every response.
This is different from word-by-word autocomplete. It proposes a whole next turn based on the conversation already in the current thread.
Who can use it
According to OpenAI’s help article, eligibility is currently narrow:
- Personal ChatGPT Pro plan
- User aged 18 or over
- Latest Codex desktop app
- Local or SSH thread
- GPT-6 Astra or GPT-6.1 Sol selected
It is not a general ChatGPT web feature. OpenAI says Work predictions may appear in local Work threads and conversations, but the main beta is described for personal Pro accounts.
Cost, context and privacy
Generating predictions does not consume Codex limits or credits during the beta. Once a user sends a message, including one accepted from a prediction, normal usage rules apply.
OpenAI says predictions use the conversation in the current thread. They do not independently look up memories or retrieve information from connected apps. Context that is already present in the thread can still influence the suggestion.
That makes review essential. A suggestion can sound plausible while missing a constraint the user has not yet stated.
How to turn Composer predictions off
The feature is on by default for eligible accounts. In the Codex desktop app, open Settings → General → Composer and switch off Show predictions.
If predictions do not appear, check the app version, plan, age eligibility, model and thread type. Long conversations may also take longer, and the system may simply choose not to show a suggestion.
Brighton Locals’ view
The manual send step is the right safeguard. Predicting “run the tests” can save friction; predicting approval for a deployment, deletion or paid action is different. Users should treat the suggestion as draft text, not as a recommended decision.
The beta will be genuinely useful only if it reduces repetitive steering without encouraging people to approve work they have not reviewed. OpenAI has not published independent accuracy figures or confirmed how pricing will work after beta.
What to watch next
Look for expansion to more plans, models and clients, along with clearer post-beta usage rules. The most important evidence will be whether users accept useful suggestions frequently without an increase in mistaken approvals.
Bottom line
Composer predictions can shorten the back-and-forth in Codex, but Tab should be treated as “put this in my draft”, never “I agree”.
Featured image: OpenAI wordmark, a public-domain text logo via Wikimedia Commons. OpenAI is a trademark of its owner.
Leave a Reply