OpenAI has taken another step in automating work with its coding assistant. Since Friday, October 9, 2026, the Codex desktop app has offered so-called “composer predictions” in beta: when Codex finishes responding, a suggestion of what the user probably wants to say next may appear in the message box. The announcement was posted that day on the company’s official developer forum.
The feature is available to users aged 18 and older on a personal ChatGPT Pro plan, in all regions where the service is offered. It only works in local and SSH-connected threads in the desktop app, and only with the GPT-6 Astra and GPT-6.1 Sol models. It is not available in web chat or on other plans, although the help center notes that it may appear in local threads of the Work environment.
The way it works is simple. After Codex responds, if the system thinks it has a good suggestion, it shows it in the message box. Pressing the Tab key adds the text to the box, where it can be reviewed or edited before sending. OpenAI stresses one point: accepting a prediction does not send the message. The final decision always rests with the user, who can ignore the suggestion and write whatever they want.
Predictions are turned on by default for eligible users. Anyone who prefers not to see them can turn them off in the app’s settings, under General and the Composer section, by switching off the “Show predictions” option. The system may also decide not to show any suggestion if it has nothing useful to offer, so one will not appear after every response.
As for cost, OpenAI explains that during the beta, generating predictions carries no extra charge, does not count toward Codex usage limits and does not consume credits. When the user sends a message, including one accepted from a prediction, the usual limits and billing apply. The company warns that suggestions can take several seconds to appear, and longer in long conversations, but nobody has to wait before starting to type.
An important aspect is what information the system uses. According to the help center, predictions are based on the conversation in the current thread. If that thread already contains context from earlier work or connected apps, it can shape the suggestion, but predictions do not look up the user’s memories or retrieve data from connected apps on their own. The feature can also be used with memory turned off.
The tool has deliberate limits. For now it suggests a complete message and does not complete, word by word, what the user is already typing. Nor can it continue a task on its own: it does not automatically accept or send follow-up messages. OpenAI recommends reviewing every prediction, because it may not match the user’s intent, and asks that problematic suggestions be reported to support with details of the model and app version.
The idea did not come out of nowhere. Predictive writing has been part of digital products for years: in 2018 Google introduced Smart Compose in Gmail to suggest phrases while writing an email, and from 2021 GitHub Copilot popularized AI code autocompletion. What is new in Codex is that the prediction applies not to code but to the conversation with the agent itself: the system tries to anticipate the next instruction the programmer would give.
That difference has implications. When working with coding agents, much of the time goes into short, repetitive messages: asking to run the tests, fix a specific error, continue with the next step or review a change. If the system gets those instructions right, the user saves time and keeps their momentum. Early reactions on social media mixed praise for the accuracy of the suggestions with complaints about limiting the feature to Pro subscribers.
It also raises questions. A tool that proposes what to say can influence, even subtly, the programmer’s decisions. If users accept suggestions without reading them carefully, human control over the agent weakens. OpenAI appears aware of that risk, which is why it designed the feature so that it never sends anything on its own and so that predictions can be easily turned off.
The launch comes during a busy week for OpenAI, which began rolling out GPT-6 in ChatGPT on October 7 and introduced its Decisions API in public beta on October 6. Codex is a central part of the company’s strategy to compete with Anthropic’s Claude Code and other assisted coding tools vying for professional developers.
The fact that the feature is reserved for the most expensive subscription suggests OpenAI wants to measure its use and cost before expanding it. Generating a prediction after every response requires additional compute, and the company has chosen not to charge for it during the beta. When that phase ends, it remains to be seen whether it stays free, is folded into normal limits or reaches other plans.
For users, the advice is to treat predictions as what they are: a useful shortcut, not an autopilot. Used well, they can reduce the friction of long conversations with a coding agent; used carelessly, they can lead people to accept steps they have not thought through. As OpenAI itself notes, a prediction may not reflect what the user wants, and responsibility for what gets sent remains human.
