On October 9, OpenAI launched a test feature in the Codex desktop app called “Composer predictions.” After Codex finishes a turn, a predicted next message appears in the input box. Press Tab and it is filled in; whether to edit it or send it is up to the user.
It is day 5 of the “one update every day for 28 days” pledge from Codex lead Thibault Sottiaux. On social media he said the predictions are accurate enough that they “Often leads to a double take,” often making people stop and look twice.
Who can use it, and how
According to OpenAI’s Help Center and a developer community announcement, the feature is currently scoped as follows:
- Users: individual ChatGPT Pro subscribers aged 18 or older, in all supported regions; during the test it is generally not available on other plans.
- Where: local threads and SSH threads in the Codex desktop app, with the model set to GPT-6 Astra or GPT-6.1 Sol. It is not available in the web version or in Chat in the desktop app.
- How it works: pressing Tab only places the text in the input box and does not send it. The announcement puts it this way: “Accepting a prediction does not send it automatically.” A suggestion also does not appear after every reply.
- Setting: it is on by default for eligible users. To turn it off, go to Settings → General → Composer → Show predictions.
On billing, generating predictions during the test costs nothing extra and does not count against Codex usage limits or credits. The message you send after accepting a prediction is counted toward usage as usual.
On data, OpenAI states that predictions use only the context already in the current thread. They do not pull in memory separately, and they do not read information from connected apps. OpenAI has not said which Pro accounts count as “eligible” or which model generates the predictions.
Same Tab key, different thing being completed
The Tab key has long been claimed by AI coding tools. Cursor’s Tab completion and GitHub Copilot’s next-edit suggestions both predict code: what to write after the cursor, or where to edit next. The user writes line by line in the editor while the tool guesses line by line alongside.
Codex puts the feature somewhere else. Its home turf has always been the conversation box: the user gives an instruction, the agent edits the repository and runs tests, then comes back with a report. After a turn, what the user is most likely to say is often a short line such as “run the tests,” “change this one too,” or “commit it.” Predicting those lines saves the back-and-forth between user and agent, not typing time.
That design also raises a boundary question. When code completion guesses wrong, the editor makes it obvious at a glance. When a conversational prediction guesses wrong, a casual “continue” or “merge it” may get sent, and the consequences land in what the agent does next. OpenAI’s help page warns that suggestions may not match the user’s intent and asks users to review them before sending; not sending automatically on Tab reflects the same concern.
Day 5 of 28
Starting October 4, Sottiaux set a rule for the team: over 28 days, ship each day either an update that clearly improves things for most Codex and ChatGPT Work users, or a full usage reset for everyone. On day 5, the item the team delivered covers only individual Pro users, which fits poorly with the pledge’s standard of being relevant to most users. No one in the community has pressed the point so far.
The announcement does not say whether the test will expand to Plus and enterprise plans, or whether predictions will start counting against usage once the test ends.
Sources: OpenAI Help Center, CocoLoop, OpenAI developer community announcement, Thibault Sottiaux’s social media posts. Eligible plans, model names, the opt-out path and billing rules follow the wording of OpenAI’s Help Center and community announcement.