Codex now writes your next message for you. Personal Pro users have it on by default.
Codex has started guessing what you were about to type and writing it into the message box for you. OpenAI calls the feature composer predictions, it is in beta, and OpenAI's help article says it is "on by default for eligible users", read at 21:35 UTC on October 10.
OpenAI's own card for the Codex quickstart, the documentation set the new beta lands in. Image: OpenAI
The mechanics are simple. After Codex finishes a response, a suggested next message may appear in the box. Press Tab and the text drops in, where you can edit it or throw it away. "Accepting a suggestion doesn't send it automatically," OpenAI's ChatGPT release notes say under an October 9, 2026 heading, read at the same time.
The eligibility list is narrow and specific
Composer predictions is "in beta for users aged 18 and older on personal ChatGPT Pro plans in all supported regions", OpenAI says, and runs only in local and SSH threads in the Codex desktop app with GPT-6 Astra and GPT-6.1 Sol. It does not appear in Chat on the web or desktop. In Work, OpenAI says predictions "may appear if you use local Work threads and conversations".
That rules out most people who use Codex through something else. OpenAI does not say which of the personal Pro tiers it reshuffled on September 29 count as eligible, only that the plan must be a personal Pro one rather than a workspace seat.
Switching it off takes four steps: open Settings in the Codex desktop app, go to General, then Composer, then turn off Show predictions. The same control turns it back on.
It is free during the beta, and so is every prediction you ignore
"During the beta, composer predictions are included at no additional cost for eligible Pro users," OpenAI says, and generating one "doesn't count toward your Codex usage limits or consume credits". Sending the message is what costs: once you hit send, with or without an accepted prediction, normal Codex usage limits and billing apply.
Predictions read the conversation in the current thread and nothing else. They "don't look up memories or retrieve information from connected apps on their own", OpenAI says, and work with memory switched off, although memory context that has found its way into the thread can shape them. Codex memories are a separate opt-in under Settings and Personalization.
OpenAI publishes no accuracy figure at all
There is no acceptance rate, no quality benchmark and no claim about how often a prediction is the message you wanted. What OpenAI does publish is a list of ways the feature will underwhelm: "the composer may choose to skip a prediction", a suggestion can take several seconds to arrive and longer in a long thread, and "a prediction won't appear after every response". That last one comes with its own reassurance that nothing is broken.
During the beta a prediction is a complete next message rather than an autocomplete that fills in as you type, and OpenAI says predictions cannot accept or send a follow-up on their own. On macOS it says the feature needs no extra keyboard permissions, and that a permission prompt claiming otherwise is worth checking.
Autocomplete arriving in the prompt box rather than the code is a small change with a large surface. OpenAI's own advice is the obvious one: "Predictions may not always match your intent, so review them before sending."
