Identify the affected access path
According to current model guidance, the date applies across ChatGPT plans, including Business, Enterprise and Edu. An application using its own API billing is a separate case. Instead of listing every system containing the word GPT, record the actual authentication path and model setting. An employee may use Codex through ChatGPT while an internal application independently uses an API key. Both need an owner, but not necessarily the same change date. For each workflow, record the surface, account type, selected model and technical contact. This avoids unnecessary changes to unaffected integrations while exposing dependencies hidden in individual employees’ working practices.
Use current guidance rather than old defaults
On 23 September, the current model page recommends GPT-6 Sol for Plus, Pro, Business, Enterprise and Edu when available. For Free and Go it names GPT-6 Luna in the desktop app, again subject to access. Earlier September guidance referred to GPT-5.6 Sol, so it should not be the only basis for the decision. Check the actual workspace and client rather than forcing one model identifier everywhere. Choose the reasoning level deliberately as well: a more intensive setting may take longer and consume more usage. The appropriate replacement is the available configuration that reliably passes your own workflow’s acceptance test.
Review saved settings and recurring tasks
The visible model picker is only part of the inventory. Inspect workspace defaults, managed configuration, custom agents, scheduled tasks and scripts. Record whether each model is explicitly selected or inherited from a default. Ask employees about important recurring tasks whose original setup they no longer remember. Give each relevant finding an owner and a test case. Keep changes traceable in a concise migration record. Avoid indiscriminate replacement across unrelated projects: a model identifier may appear in historical notes or API examples without representing a live workflow that needs changing. The aim is to find operational dependencies, not simply to eliminate a string from every file.
Example: a weekly sales report
Suppose a team produces a weekly report from approved sales exports. Before switching, rerun a previously reviewed period with the existing workflow and preserve the reference result. Then use the replacement with identical data, definitions and instructions. Check totals, missing values, sources and output format. Different wording is not automatically a failure; a changed figure without explanation is. The scheduled trigger and delivery location also need to continue working. This is a proposed testing method. Assign someone to inspect the first real report after migration before it is circulated, rather than treating a successful interactive trial as proof that the unattended workflow is ready.
Switch before the deadline and keep a fallback process
Schedule the switch sufficiently before 14 October to expose access problems or quality failures while there is time to respond. After comparison tests pass, move a few tasks first. Record the model, reasoning level, review result and date. A fallback cannot rely indefinitely on a retiring model; prepare a manual process or another verified available configuration. Check the next scheduled execution and its usage afterwards. Migration is complete when the affected workflows operate with the new settings and their owners know what to do if a run fails, not merely when the model name in a configuration has been changed.
Common questions about model retirement
Must API applications switch by 14 October? This announced deadline does not affect the OpenAI API. Review API lifecycle notices separately. Will existing tasks automatically behave identically? Do not assume so; model behaviour and available options can differ. Is the newest model always the best replacement? Access, quality, runtime and the total cost of your task determine the choice. What if I do not have administrative rights? Report the affected workflow, model identifier and owner rather than bypassing local policy. What is the most useful first step? An inventory of important tasks and a small comparison test provide more clarity than rewriting every prompt without a specific problem to solve.
Keep it verifiable
Primary sources
- OpenAI: Work and Codex models and GPT-5.5 retirementSource checked:
- OpenAI: ChatGPT and Codex changelogSource checked:
- OpenAI: API changelogSource checked:
