OpenAI Adds Next-Message Predictions to Codex Desktop Beta
The optional feature proposes a complete follow-up after a response for eligible Pro users, while leaving review and sending under the user's control.
Edited by Tyronne Panaino
OpenAI added Composer predictions to the Codex desktop app in beta on October 9. The optional feature can propose a complete next message after Codex responds, using the conversation in the current thread as context. OpenAI's ChatGPT release notes say the beta is available to adults on personal ChatGPT Pro plans in supported regions.
The change matters to developers who work through long local or remote coding conversations because it places a possible follow-up directly in the composer. It does not send anything automatically. The user can accept the text with Tab, edit it, ignore it or write a different message, according to OpenAI's Composer predictions guide.
A suggestion appears only after a Codex response
Composer predictions work at the turn boundary. After Codex finishes a response, a suggested follow-up may appear in the message box. The feature may also skip a prediction, so an empty composer does not necessarily indicate a fault. A user can begin typing without waiting for the suggestion.
Pressing Tab places the proposed text in the composer. The user still reviews the wording and deliberately sends the message. OpenAI describes the output as a complete next message rather than a word-by-word completion for text that is already being typed. That distinction makes this closer to a suggested conversational move than conventional inline autocomplete.
The design preserves a visible checkpoint between prediction and action. A poor suggestion cannot continue a task by itself, and accepting one does not trigger the send operation. OpenAI also warns that suggestions may not match the user's intent, making review part of the documented workflow rather than an optional safety gloss.
Access is limited by plan, client and model
The beta has a narrower footprint than Codex as a whole. OpenAI lists personal ChatGPT Pro accounts for users aged 18 and older, the latest Codex desktop app, and local or SSH threads. The supported models are GPT-6 Astra and GPT-6.1 Sol. The feature guide says predictions are not available in Chat, so this should not be read as a general ChatGPT composer change.
Those limits affect how readers should interpret the announcement. This is not a broad rollout across every Codex client, model or account type. It is an early desktop interaction for a specified set of paid personal users. The source pages do not provide a general-availability date or promise expansion to more plans and models.
For eligible users, predictions are enabled by default. The control sits under General and then Composer in Codex settings, where Show predictions can be turned off and later enabled again. That setting gives users a direct way to remove the extra text from their workflow without changing the model or abandoning the thread.
Current-thread context defines the prediction boundary
OpenAI says the feature uses the conversation in the current Codex thread. It does not independently look up memories or retrieve information from connected apps. Context that is already present in the thread can still influence a suggestion, including information brought in through earlier work. Predictions can also be used with memory turned off.
This boundary is important because a plausible follow-up may reflect more than the immediately preceding answer. In a long coding session, earlier requirements, tool results and discussion already present in the thread can shape the proposed message. The source does not claim that the prediction is a summary of the full thread or that it will consistently identify the best next engineering action.
OpenAI says generating a prediction during the beta does not consume Codex usage limits or credits. Normal usage rules and billing apply once a message is sent, including when the user sends an accepted prediction. The feature therefore changes how a message can be drafted, not how an accepted request is processed after submission.
The next evidence is quality and wider availability
The immediate checkpoint is whether OpenAI expands Composer predictions beyond the current personal-Pro, desktop, local-or-SSH and two-model envelope. Other useful signals would include documented changes to the context boundary, controls for managed workspaces, and evidence about how often suggestions are useful rather than distracting.
Neither official page reports acceptance rates, productivity gains, error rates or independent testing. It would be premature to treat the beta as proof that predicted follow-ups improve software work. The confirmed change is more specific: eligible Codex desktop users can receive a proposed next message, and the user remains responsible for reviewing and sending it.
Status
Confirmed. Internal confidence is medium because both sources are official OpenAI Help Center pages describing the same first-party beta, without independent evaluation of quality or workflow impact.
Sources
Update note: Last reviewed October 10, 2026. We will revise this article if OpenAI changes eligibility, supported models, context use, usage accounting or rollout status.
Sources
- OpenAI — ChatGPT release notes — official
- OpenAI — Composer predictions in Codex — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.