A missing suggestion in Codex is not, by itself, a reason to interrupt your work. OpenAI's new composer predictions beta can help prepare a follow-up message, but it is optional assistance rather than a required stage of every conversation. The useful distinction is between a feature you cannot currently access, a suggestion that simply does not appear, and a suggestion you should rewrite.
OpenAI announced the beta in its October 9, 2026 release notes. It is available to adults on personal ChatGPT Pro plans in supported regions, using the latest Codex desktop app. The listed combination is a local or SSH thread with GPT-6 Astra or GPT-6.1 Sol. Predictions are enabled by default for eligible users. OpenAI's October 9 release notes
That narrow scope matters. An announcement about Codex does not mean every place you use ChatGPT gains the same feature. Before treating an absent prediction as a technical fault, establish which experience you are actually using.
Why Codex composer predictions are not showing
OpenAI says predictions can be skipped. They may take several seconds, with longer waits in lengthy conversations. Its recommended checks are the app version, account, selected model, thread type and the Show predictions setting under Settings, General, Composer. Ordinary Chat on web or desktop is outside the documented beta. You can keep typing without waiting. Composer predictions documentation
These observations call for different responses:
| What you observe | What it establishes | A proportionate response |
|---|---|---|
| Your setup falls outside the published eligibility list | The announcement does not promise access in that setup | Continue manually and check the current scope before changing plans |
| An eligible conversation has no suggestion after a response | One optional suggestion is absent | Write the next instruction yourself rather than interrupt a useful task |
| Suggestions repeatedly fail to appear in an eligible, updated setup | You have a repeatable observation worth investigating | Record the environment and timing, then report the specific behavior |
| A suggestion appears but proposes the wrong next step | The wording needs your judgment | Rewrite it around the outcome you actually want |
This table is an editorial interpretation of the documented behavior, not a diagnostic test or a guarantee that a particular account is working correctly.
Accepting a suggestion is separate from sending it
Pressing Tab places the proposed message into the composer. It does not submit it. During the beta, generating predictions does not consume Codex usage or credits, while messages you send follow normal usage limits and billing. OpenAI release notes
For a team, that makes the composer a useful review point. A follow-up can be grammatically clear while still aiming at the wrong result. Read the action it requests and the evidence that would count as completion, rather than judging only whether it sounds natural.
Consider a hypothetical product review. You asked an assistant to compare two draft landing pages, but the proposed continuation asks it to publish the preferred page. The comparison and the publication are different decisions. A better continuation would identify the chosen draft and request the specific revision or review you intended. This is an illustrative example, not a report of observed Codex behavior.
Context can explain a suggestion without validating it
Predictions use information already in the current thread. They do not independently retrieve memories or connected-app information, and memory does not need to be enabled. They propose a complete message rather than updating word by word as you type. OpenAI's feature documentation
Our practical implication: when the project has changed direction, state the new objective clearly in the conversation. A smooth continuation of an old plan can still be a poor next step. The same applies to any draft instruction, whether it came from an assistant, a colleague or your own earlier notes.
What to do with the beta now
For eligible users who already work in Codex desktop, predictions are worth evaluating against actual work: did the suggestion preserve your intended scope, and how much rewriting did it need? Keep a few examples of useful and unhelpful continuations instead of counting accepted suggestions as evidence of productivity.
If you suspect a persistent issue, OpenAI asks support reports to include the suggested text, expected behavior, model and desktop app version; remove sensitive information from screenshots. You can turn predictions off using Show predictions. OpenAI support guidance
For a clearer process around the work itself, use NEXAIUM's three-step AI workflow. A defined research, draft and reuse sequence gives you a better basis for choosing the next instruction, whether a suggestion appears or not.
Product scope verified October 13, 2026. This article explains the October 9 beta announcement; eligibility and behavior may change. It does not claim hands-on feature testing or measured productivity gains.