OpenAI sets October 14, 2026 as the last day GPT-5.5 runs a scheduled task

The substitution is small. In Codex signed in with a ChatGPT account, gpt-5.5 becomes gpt-5.6-sol ^[3]. Anything calling the OpenAI API keeps the model string it has, because OpenAI says the API is exempt ^[4]. The work is finding the tasks.

That happens in Scheduled, which lists active, paused and completed tasks along with recent runs ^[6]. Neither the Codex CLI nor the IDE extension has that view ^[7][8]. Of the four clients the page describes, two show the list ^[21]. The CLI is where the docs suggest you prepare and test a prompt, skill or script in the first place ^[26].

Paused tasks sit in the same view as active ones ^[6]. A task paused in March and pinned to GPT-5.5 will still be there on October 15, and the page does not say what a run does after the retirement date ^[22].

Some tasks fire on events. On eligible plans, a task can run when a new Gmail message arrives, when a message lands in a selected Slack channel, or on pull request activity in a repository ^[13]. One task can carry several event triggers, and it cannot combine event triggers with a time-based schedule ^[14]. Event triggers work in ChatGPT on the web and on mobile, not in the desktop app, the CLI or the IDE extension ^[13]. In a managed workspace they exist only where an administrator has granted the Allow event-triggered scheduled tasks permission ^[16]. That permission is also the place to look when you are counting how many of these a team could have created.

Then the local ones. In the desktop app a scheduled task can run against a local project, either in the project directory or on a new worktree, which keeps its changes separate from unfinished local work ^[9][25]. The requirements are ordinary: the machine powered on, the app running, the project still on disk when the run is due ^[10].

Editing a task in place is cheaper than recreating it. A standalone task starts each run from its saved prompt, while a task created inside a chat returns to that chat with its existing context ^[17]. A rebuilt event-driven task has its preconditions to satisfy again: @ChatGPT added to every Slack channel it watches, and repository access for the connected GitHub app ^[15]. Web runs start clean, with no local folder or worktree carried over, so durable instructions belong in the task prompt or an attached skill ^[11][12].

Once the model changes, the page's own advice for a new task applies to an edited one: test the prompt in a regular web chat, then review the first few runs and adjust the prompt, tools or cadence ^[18].