Codex can suggest your next message. Tab accepts it; it does not send. The post is a community note, not a newsroom launch

What did OpenAI publish on 9 October?

The dated object is a Codex-category post on community.openai.com, created 9 October 2026 at 20:01 UTC, signed @sps. The lede is that composer predictions are now in beta for personal ChatGPT Pro users aged 18+ in the Codex desktop app. After Codex responds, “it can suggest a follow-up based on the current conversation.”

The same post’s detail block is four lines. Where: local and SSH threads with GPT-6 Astra or GPT-6.1 Sol, in all supported regions. How: press Tab to accept, review or edit, then send; “Accepting a prediction does not send it automatically.” Control: enabled by default; turn it off under Settings → General → Composer → Show predictions. Cost: during beta, generating predictions “costs nothing extra and uses no Codex limits or credits.” Sending still follows normal usage and billing.

We searched openai.com/news for a matching item and did not open one. This is a desktop-composer change, not the 7 October GPT-6 Chat-tab rollout. It is also not Together Link’s Codex launcher, which points Codex at Together-hosted models.

Where does it work, and where does the Help page say it does not?

The Help Center article “Composer predictions in Codex” restates the beta, the 18+ personal ChatGPT Pro limit, the Codex desktop app, local and SSH threads, and the two supported models. “Select one of these models in your local or SSH thread to use the feature.”

Two Help sentences sit next to each other. One says predictions “are not available in Chat on web or Desktop.” The next says “In Work, predictions may appear if you use local Work threads and conversations.” A later FAQ says that during the beta, predictions “are available only on personal ChatGPT Pro plans in the Codex desktop app” and “aren’t available on other plans on the web, or in Chat.” Those are the page’s sentences. We are not collapsing them into one eligibility rule.

Predictions can appear in existing conversations that meet the eligibility lines. They “may take longer to appear in long conversations.” A prediction “won’t appear after every response,” and “the composer may choose to skip a prediction.” That skip is documented behaviour, not a fault we observed.

How do you accept a suggestion, and what does Tab not do?

The Help page’s steps are: update the Codex desktop app, sign in to a personal ChatGPT Pro account, open a local or SSH thread with a supported model, send a message, then look for a suggestion in the message box after Codex finishes. Press Tab to accept. Review or edit. Send when you are ready.

Accepting “adds text to the message box — it does not send the message for you.” You can start typing your own message without waiting. During the beta, predictions “suggest a complete next message. They don’t update word by word as you type.” They also “don’t automatically accept or send follow-up messages.”

That last line matters if you are comparing this to an agent that continues a task on its own. Composer predictions are a suggested user utterance, not a run. For agents that do act through a shell, the control problem is closer to computer-use agents and to a backdoored model driving Codex CLI than to a greyed-in next message.

What context do predictions use, and what do they cost?

Both pages say predictions use context already in the current thread. The community post: they “don’t independently retrieve memories or information from connected apps.” The Help page: if past work or connected-app context is already in the thread, “it can inform a prediction,” but predictions “don’t look up memories or retrieve information from connected apps on their own.” You can use them with memory off. Memory already in the thread “can personalize predictions.”

During the beta, generating predictions is included at no additional cost for eligible Pro users and “doesn’t count toward your Codex usage limits or consume credits.” When you send a message, including an accepted prediction, “normal Codex usage limits and billing apply.” Predictions “can take several seconds to appear after Codex finishes responding, and longer in long conversations.”

Messages you send “follow Codex’s retention and deletion policies.” The Help page points at Data controls in ChatGPT and “How your data is used to improve model performance.” We did not open those secondary policy pages in this pass.

How is this different from the new-conversation chips?

The Help FAQ draws a clean line. “Suggestions on the new conversation page help you start a task. Composer predictions suggest your next message in an existing conversation, after Codex finishes responding. Show predictions controls composer predictions.” Turning the setting off does not, on that page, remove the empty-thread starter chips.

The feature is on by default for eligible users. Settings → General → Composer → Show predictions is the documented off switch. The Help page says composer predictions “don’t require additional keyboard permissions” on macOS. We did not click that toggle. If you are choosing a coding assistant rather than a composer pref, start from how to choose an AI assistant or from JetBrains Mellum 2.1, which is an in-IDE model, not this desktop suggestion.

What did we not test?

We have not opened the Codex desktop app, signed in, accepted a prediction, or timed how long a suggestion takes. We did not compare Astra against GPT-6.1 Sol. We did not reproduce Work-thread behaviour. Direct HTTP fetches of the Help URL returned 403; the Help wording above is from the converted page we opened through the fetch tool on 10 October. Secondary social posts that quote internal “most loved” lines are not on the two pages we used as primary sources, and we are not repeating those lines.