GPT-5.5 Retirement in Codex: What to Change

GPT-5.5 retirement in Codex is effective for users signed in with ChatGPT from October 14, 2026, at noon PDT, or 19:00 UTC, according to OpenAI’s newly dated notice. The OpenAI API and API-key-authenticated Codex sessions are unaffected. The immediate task is to find affected saved selections, not replace every reference across both authentication routes.

GPT-5.5 retirement in Codex: the new notice’s scope

The October 14 entry in the official changelog names workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks and scripts as places to update. This is an effective notice, rather than our inference that a previously announced deadline has passed. We have not tested enforcement across individual accounts or clients.

The current Models guidance recommends GPT-6.1 Sol, using the exact ID gpt-6.1-sol, for Plus, Pro, Business, Enterprise and Edu when available. For Free and Go, it recommends GPT-6 Luna in the desktop app when available. The site’s GPT-6.1 Sol page and Luna guide provide separate model and usage context.

The Work and Codex pricing page, checked October 15, also carries the scoped retirement notice and no longer lists GPT-5.5 in its Standard credit table. That table observation is not an API shutdown or a newly announced price for the remaining models.

Classify the caller before editing the selection

Consider our hypothetical mixed-authentication team. One developer uses a ChatGPT-signed-in desktop session. A scheduled internal job runs Codex with the team’s API key. Both refer to gpt-5.5, but the notice puts them on different sides of the retirement boundary.

Our suggested inventory records the caller, sign-in route, selected model, client and owner. Mark the developer’s saved choice for migration. Mark the API-key job as unaffected by this particular retirement, rather than automatically changing it alongside the desktop session. An unrelated decision to migrate that job would need its own reason and checks.

The same executable name does not establish the same authentication route. Ask the job owner which route the actual run uses and record that answer without copying credentials into the inventory. If the route is unknown, resolve it before assigning a replacement. A guessed route can produce an unnecessary edit or leave an affected task unchanged.

Keep source-code examples and active callers distinct as well. A historical note quoting an old model ID is not necessarily an executable selection. Label that reference as documentation, an example, or an active configuration before deciding whether it needs a change. This makes the inventory actionable instead of turning a text search into an indiscriminate replacement.

Availability and a default are separate controls

OpenAI’s workspace model availability guidance says GPT-6.1 Sol is off by default in Enterprise and Edu until an administrator enables it. Changing local configuration does not grant access. The guide distinguishes product surfaces and authentication boundaries; an administrator should confirm the replacement where affected users actually work.

In our hypothetical team, an administrator can choose a preferred replacement while a developer still cannot select it in the relevant client. That result calls for an availability check, not a second spelling of the same model ID. Confirm the intended identity, workspace and client before changing more settings.

Separate the person who approves model access from the owner who updates a task’s stored selection. They may be the same person, but the two decisions solve different problems. A migration record should show both that the intended user can use the replacement and that the affected task now asks for it.

Do not broaden a model-access fix into a permissions change. If the task needs a different sandbox, a network destination or access to another system, that is a separate request. The workspace guide explicitly distinguishes model availability from what a local run is permitted to do.

Trace one task through its stored entry points

Our next example is an affected weekly review task. The operator changes an interactive model selection but leaves the stored task configured for gpt-5.5. The next scheduled invocation therefore remains a migration item even though the operator’s new chat uses a different model.

Our recommendation is to follow that one task from its saved definition to the command or client that actually launches it. Record every explicit model selection encountered and the owner responsible for it. Do not assume an interactive change rewrote the saved definition, or invent a configuration-precedence rule that the relevant client has not documented.

For a second example, a custom agent template can create new tasks with an old selection. Updating one existing task would not resolve the template’s future output. Review the template and a newly created task separately. Conversely, changing the template alone is not evidence that every existing task has been migrated.

Use the control that belongs to the affected surface

The Models page documents three local choices: the desktop composer control, interactive CLI /model, and an explicit –model or -m launch option, including non-interactive runs. It also documents a shared config.toml for desktop, CLI and IDE extension, with a model entry for the local default.

Apply that distinction to a developer who changed the desktop selection but also launches a named command from the terminal. Our suggested check examines that command’s explicit selection separately from the saved local default. Do not infer a precedence order or assume that changing one control rewrote another; inspect the configuration relevant to each entry point.

A second local example concerns shared configuration. If the operator edits the model entry intending to migrate the CLI, review the desktop and IDE uses of that same file as part of the change’s scope. Conversely, a successful selection in one existing conversation is not evidence that the shared default has been edited. These checks distinguish a session choice from a configuration change without inventing automatic propagation.

The workspace guide states that Codex cloud does not support changing its default model. A hosted-task owner should therefore not follow a local-file recipe as though it were an available cloud-default control. Check the supported model for the hosted workflow and the signed-in identity; surface an unavailable required option rather than repeatedly editing a local file that the cloud-default guidance does not describe.

For a separate API migration deadline, see our GPT-5.3-Codex API deprecation report.