Codex can now draft the next message you're about to send. OpenAI calls the feature composer predictions. As of October 10 it's in beta for personal ChatGPT Pro users on the Codex desktop app, and OpenAI Developers said the suggestions come from the conversation and from how you usually talk to it — one of the most loved new features the team has tested internally.
Tibo (@thsottiaux) posted it at 03:47 Beijing time as day 5 of the 28-day Codex plan. He says the predictions are often surprisingly on target, and that they come with Pro plans without consuming usage. OpenAI's help page, which went up the same day, fills in the rest: the feature works in local and SSH threads with GPT-6 Astra or GPT-6.1 Sol, and during the beta, generating a prediction doesn't count toward Codex usage limits or consume credits. Sending the message still does, including if you accepted a suggestion.
It isn't available on Plus, Business or other non-Pro plans, and it isn't in the CLI or the IDE extension — only in the desktop app. Press Tab to drop a suggestion into the box, review or edit it, then send; accepting doesn't send it for you.
A Windows sandbox update that arrived the same night
Separately, OpenAI Developers announced a new Windows sandbox built on Microsoft's Execution Containers, or MXC. The post says it aims for faster setup, stronger network enforcement and more granular file access, and that it needs a compatible Windows 11 device.
OpenAI's Windows sandbox docs go a bit further. MXC is now the recommended sandbox on devices that support it, and it can run without administrator-approved setup, extra Windows accounts or local firewall rules. The docs say Microsoft added the needed process isolation in Windows 11 24H2 (build 26100.9278) and 25H2 (build 26200.9278). On machines that don't support MXC, Codex can fall back to the older elevated or unelevated sandboxes.
For how Pro usage and other Codex caps work, see our guide "OpenAI Keeps Codex's Weekly Caps Vague: What Each ChatGPT Plan Actually Gets".
