OpenAI Speeds Up Mid-Task Steering in Codex Desktop
The ChatGPT desktop update makes follow-up direction reach a running Codex task sooner while preserving a setting that can defer messages to the next run.
Edited by Tyronne Panaino
OpenAI began rolling out faster steering for Codex in the ChatGPT desktop app on October 8. A follow-up sent while a task is running can now reach Codex sooner, giving a developer a faster way to correct an approach, add missing information or change direction before the run finishes.
The official ChatGPT release note also identifies a control for deciding whether follow-up messages steer the current run or wait for the next one. The change matters for long-running coding work because not every new instruction has the same urgency: some should redirect active work, while others are better treated as the start of a later run.
Steering now has a clearer timing choice
The supported product change is narrow. OpenAI says Codex can respond sooner when a user sends a follow-up to a running task. That makes steering a live interaction rather than requiring the user to wait for the current run to complete before supplying a correction.
The release note does not describe a new model, coding capability or execution environment. It is an interaction and task-routing update inside the desktop app. Existing articles about Codex Cloud, remote control and iOS task links address where tasks run or how people reach them; this release addresses when a follow-up is applied to active work.
Follow-up behavior separates redirection from queueing
OpenAI places the choice under Settings, General, Follow-up behavior. One option lets a follow-up steer the current run; the other waits for the next run. That distinction can help users avoid unintentionally interrupting an approach that should finish, while still making an urgent correction available during execution.
The note establishes that the setting exists, but it does not explain how Codex reconciles a late correction with work already completed during the run. It also does not say whether every tool call can be redirected at the same point or whether a follow-up changes approval requirements. Those behaviors should be verified in the actual task and workspace rather than inferred from the announcement.
Faster is confirmed but not quantified
OpenAI provides no measured latency, earlier baseline, completion-rate comparison or independent reliability test for the faster response. The note also omits a plan list, operating-system matrix and geographic rollout boundary. Users may therefore see the setting at different times, and the announcement alone does not prove a specific speed improvement on every task.
The next useful checkpoint is documentation that defines client availability and shows how steering behaves across different stages of a running task. Measured evidence would also be needed to determine whether faster steering reduces wasted work or improves final output quality rather than only shortening the time before Codex acknowledges a change.
Status
Confirmed. OpenAI's dated release notes establish the faster-steering rollout and the current-run-versus-next-run control. Internal confidence is medium because the evidence is one first-party product note without quantified or independent operational verification.
Sources
Update note: Last reviewed 2026-10-09. We will revise this post if OpenAI publishes client availability, measured steering performance or more detailed task-handling behavior.
Sources
- OpenAI — ChatGPT release notes — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.