Codex composer predictions suggest your next message
OpenAI has documented Composer predictions, a beta that proposes a complete next message after Codex responds. Adults on personal ChatGPT Pro can use it in desktop local or SSH threads with GPT 6 Astra or GPT 6.1 Sol.
Press Tab to insert a suggestion, then review, edit and send it. Nothing is sent automatically. Enabled by default. Disable Show predictions under Settings, General, Composer. Suggestions may be delayed or absent, and users can keep typing. Keep the desktop app updated.
The beta covers supported regions and may appear in local Work conversations. Regular Chat on desktop and web is excluded, according to OpenAI’s guide.
Check the task before sending
The useful question is whether the proposed instruction still matches the job. OpenAI’s prompting guidance says a coding request should identify the intended behavior, relevant code or reproduction steps, important constraints and a way to verify the change. Those criteria provide a practical review checklist for any drafted instruction.
For example, a developer investigating a failed test may want an explanation before approving a code change. A message that jumps straight to editing changes the task. Check the files, requested action and verification step against what you actually want completed.
There is also a separate execution boundary. OpenAI’s security documentation distinguishes the sandbox, which limits technical access to files and networks, from the approval policy, which determines when an action needs permission. Reviewing the wording of a prompt and controlling the agent’s access address different risks. A well phrased request still needs appropriate execution permissions.
Memory and privacy have separate controls
Predictions use the current thread, including memory or connected app information already present. They perform no independent retrieval, and memory can remain off.
OpenAI’s memory documentation describes a separate local store for local Codex clients. Settings under Personalization control whether memories are enabled for the selected machine, while the /memories command controls whether a particular conversation can use or contribute to them. Required team instructions belong in project documentation or AGENTS.md. OpenAI cautions against making remembered context the only place for rules that must always apply.
Model training is another setting. On personal plans, Improve the model for everyone applies to Codex tasks as well as ChatGPT conversations. Turning it off prevents new conversations and tasks from being used for training. Codex also has a separate Include environments setting for additional environment context. Changing the ChatGPT training preference does not change that control.
Saved conversation data needs its own consideration. Codex’s retention guidance says archived conversations stay stored and may retain connected app information. Disconnecting the app prevents future access without removing those earlier conversations. Deletion removes the chat from view immediately and normally schedules permanent removal within 30 days, with exceptions for legal or security obligations and content already detached from the account for model improvement.
Free suggestions still lead to metered work
Generating predictions costs no extra money, credits or quota during beta. Sending one still incurs normal usage.
That usage depends on the work requested. OpenAI’s Codex plan guide lists the model, task complexity, context, reasoning, speed and tools among the factors affecting consumption. It directs users approaching a limit to Settings or their usage dashboard to check the exhausted allowance, credit balance and reset time. A convenient next instruction can therefore start a substantial task, so reviewing its scope is also a cost decision.
