OpenAI has added Composer Predictions to Codex in beta. The feature can suggest a next message from the current conversation after Codex responds. It is a small interaction change, but it is relevant to quality teams: a prediction can speed up repetitive follow-ups while still leaving the user in control of whether to edit, accept, or send anything.
What OpenAI announced
According to OpenAI’s October 15, 2026 ChatGPT release notes, Composer Predictions is available in beta to personal ChatGPT Pro users aged 18 or older in supported regions, using the latest Codex desktop app. It supports local and SSH threads with GPT-6 Astra or GPT-6.1 Sol.
After Codex answers, a proposed next message may appear in the composer. Pressing Tab accepts the suggestion into the message box, where it can be reviewed or edited. Crucially, accepting a prediction does not send it automatically. OpenAI says the feature is on by default during the beta and can be turned off in General settings under Composer.
Why this matters for QA engineers
For test engineers, the distinction between suggest, accept, and send is more than interface polish. It is an approval boundary in an AI-assisted workflow. A suggested follow-up could help maintain momentum in a bug investigation or test-failure triage, but it must not become an unobserved instruction that changes a repository, invokes a tool, or weakens a test assertion.
- Test the state transition: verify the suggestion appears only after a response, Tab inserts it, editing changes only the draft, and sending remains a deliberate separate action.
- Probe risky context: use threads involving production-like URLs, credentials placeholders, destructive commands, and conflicting instructions. Confirm the suggestion is reviewable and does not execute by itself.
- Check thread parity: run the same scenarios in local and SSH threads on each supported model, noting availability and wording differences.
- Capture evidence: record the original assistant response, suggestion, final edited prompt, tool approvals, and resulting diff or test output.
- Validate the off switch: disable Composer Predictions and restart or reopen a thread to confirm no suggestion remains.
A practical adoption approach
Start with low-risk tasks such as asking Codex to summarize a failed run, draft a reproducible bug report, or propose the next verification command. Measure whether suggestions reduce repetitive typing without increasing unwanted sends, invalid assumptions, or missed review steps. Keep normal code-review, test-execution, and approval gates in place; this beta changes message composition, not the need for evidence.
Because this is a beta rollout, eligibility, regions, app versions, models, and workspace configuration can affect what users see. Teams should document their observed configuration alongside results instead of assuming universal availability.