GPT-5.5 Leaves ChatGPT and Codex on October 14: Check Your Restaurant Automations First
Find affected model selections, record replacements and verify manual tests and actual scheduled runs. Includes a fictional restaurant automation register.
Ryan Speier
Quick answer: OpenAI's ChatGPT and Codex documentation, in a changelog entry dated September 14, 2026, says GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on October 14, 2026, on every plan, including consumer, Business, Enterprise and Edu. The OpenAI API is not affected. If any restaurant automation of yours actually selects gpt-5.5 as its model through a ChatGPT sign-in, such as a scheduled task, a custom agent, a workspace default, a Codex configuration file or a script, you need to change that selection to an available model and test it before that date. The model name appearing in prompt text, or in a setting on the API side, is not an affected selection. For Codex used with a ChatGPT sign-in, the documented replacement is gpt-5.6-sol (GPT-5.6 Sol). Anything a vendor built on the OpenAI API with its own API key is outside this notice, though it is worth asking the vendor which path it uses.
Who this is for
Owners and managers who set up ChatGPT scheduled tasks, custom agents or small Codex scripts for things like order drafts, weekly labor summaries or SOP lookups. If you only chat with ChatGPT by hand, check the available models and any saved preferences before your next use; the operational checklist below is mainly for recurring workflows. If your only AI use is inside a vendor's product, this notice is the vendor's problem, not yours, as long as it runs on the API.
What the notice says and what it does not say
| The documentation says | The documentation does not say |
|---|---|
| GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on October 14, 2026, on all plans. | What happens to a scheduled task, agent or script that still selects gpt-5.5 after that date. Do not assume it switches itself. |
| The retirement does not apply to the OpenAI API. | Which path a given vendor integration uses. You have to ask. |
| Replace gpt-5.5 in workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks and scripts that select a model. | A single named replacement for ChatGPT scheduled tasks. The scheduled-tasks page says to "choose an available replacement." |
| For Codex with ChatGPT sign-in, switch gpt-5.5 to gpt-5.6-sol before October 14. | Anything about custom GPTs. That is a separate topic and is not covered here. |
| Scheduled tasks can leave the model and reasoning effort on defaults or set them explicitly; tasks and their runs are listed under Scheduled in the ChatGPT desktop app, and web tasks are managed from Scheduled on ChatGPT web. | How long the check takes. That depends on how many automations you have and whether anyone wrote them down. |
Step 1: write down every automation before you touch any of them
Bayside Tacos is an invented two-location taqueria. Its manager built a few automations after reading earlier posts here. This is what the register looked like once they wrote it all down. Yours will be shorter or longer; the point is that nothing is missing.
| Automation and what it does | Where it runs and how it signs in | Model selection | Owner, last good run | Action |
|---|---|---|---|---|
| Monday par-order draft: turns Sunday's count sheet into an order draft against pars | ChatGPT web scheduled task; ChatGPT Business | GPT-5.5 selected in the task | Kitchen manager, Sept 14 | Change the selection, test, verify the next scheduled run |
| Weekly labor summary: reads the exported hours file in a local folder | ChatGPT desktop app scheduled task; ChatGPT Plus | Default, nothing selected | Owner, Sept 15 | Confirm it is on default; record it |
| SOP question helper: answers staff questions from approved SOP PDFs and cites the page | Custom agent; ChatGPT Business | gpt-5.5 selected in the agent's configuration | General manager, Sept 12 | Change the selection, test, verify the next real use |
| Shift text list: script that turns the schedule spreadsheet into a text list for staff | Codex CLI on the office laptop; ChatGPT sign-in | gpt-5.5 in the script's --model flag and in config.toml | Owner, Sept 13 | Change to gpt-5.6-sol, test, record |
| Prep checklist prompt: a saved prompt whose text says "act like GPT-5.5 did last time" | ChatGPT web, run by hand; ChatGPT Plus | None; the name is only in the prompt text | Sous chef, Sept 16 | Nothing to migrate; tidy the wording if you like |
| Online-ordering chat widget: answers menu questions on the website | Vendor's system; vendor's API key | Unknown, API side | Vendor | Ask the vendor which path it uses; nothing else from this notice |
A blank version you can paste into a note or spreadsheet:
AUTOMATION REGISTER (one line per automation) Name | What it does | Where it runs (ChatGPT web / desktop app / custom agent / Codex CLI / vendor) | Sign-in (ChatGPT plan or API key) | Model selection (default, or the exact model selected; not names in prompt text) | Owner | Last good run (date) | Replacement chosen | Tested on (date) | Next review (date)
Step 2: find where the model name is actually set
| Surface | Where to look | What to record |
|---|---|---|
| ChatGPT scheduled tasks | Open Scheduled in the ChatGPT desktop app sidebar, or manage runs from Scheduled on ChatGPT web. Open each task's settings. | Whether the model is on default or explicitly selected, and which model. |
| Custom agents | Depends on the surface where the agent was built. Open its configuration and look for a model selection; the pickers differ by surface and some may not expose one. | The model it selects, if any, and where that setting lives. |
| Workspace defaults and managed configurations | Admin settings for Business, Enterprise or Edu workspaces. The models page points administrators to "workspace model availability." | The default model your workspace hands to new chats and tasks. |
| Codex CLI, desktop app and IDE extension | They share one config.toml file. The documentation shows the setting as a line like model = "gpt-5.6". If there is no model line, Codex uses a recommended model. | The model line, or "none set." |
| Scripts and shortcuts | Any script that runs codex exec --model, and any shortcut that sets a model. | Every place gpt-5.5 is selected. A mention inside prompt text is not a selection. |
| Codex cloud chats | The models page says you currently cannot change the default model for Codex cloud chats. | Nothing to change; note it in the register. |
For the script and config side, search only the configuration file and the folders where your scripts live. Do not search session history or credential files, and keep the output to yourself; it can contain file paths and other details you would not want in a shared chat. Replace the paths with your own.
grep -n "model" /path/to/your/config.toml grep -rn "gpt-5.5" /path/to/your/scripts
Explicitly choosing a model is not a mistake. Pinning controls which model runs; it does not promise identical output from run to run, so the test in Step 3 still matters. The mistake is pinning without writing it down. Once the pin is in the register with an owner and a review date, a retirement notice is a small job instead of a surprise.
Step 3: choose the replacement and test it once
For Codex with a ChatGPT sign-in, use gpt-5.6-sol; that is the documented replacement. For ChatGPT scheduled tasks and custom agents, choose from the models the task or agent settings offer, or leave the setting on default and record "default" plus the date. This post does not recommend a specific ChatGPT model because the documentation does not, and the right choice depends on your plan and the job.
OpenAI's own advice for scheduled tasks is to test the prompt manually in a regular chat first, confirm that the selected or default model, reasoning effort and tools behave as expected, and confirm the output is reviewable. For a restaurant task, "behaves as expected" means checking the parts that cost money when they are wrong. Bayside Tacos' Monday par-order draft is deliberately a toy: five dry-goods items, whole-case pars, nothing on order, nothing perishable. A real order draft must also account for quantities already on order, usable versus damaged stock, pack rounding and shelf life, which is why it stays a draft a person approves. The test:
| Check | Pass looks like |
|---|---|
| Run the task by hand with last Sunday's count sheet, the same file the September 14 run used | An order draft comes back in the same layout the kitchen manager expects |
| Compare it with the September 14 output, line by line | Same items, same units; every difference is explained by the count, not by the model |
| Unit and pack checks | Cases stay cases, each stays each; nothing converted silently |
| The par rule | Items at or above par are not ordered; items below par are ordered up to par, not above |
| Anything invented | No item appears that is not on the count sheet |
| Reviewability | A person can read, approve or correct the draft; it is a draft, not an order, and nothing is sent to a supplier automatically |
| The next real scheduled run | After the manual test passes, watch the next actual scheduled run in the same account, environment and tool permissions. Check its timestamp and output in Scheduled; a manual test is not proof that the scheduled run works |
If a check fails, keep the September 14 output as the reference, adjust the prompt, and test again. If it still fails, pause the task and run the job by hand until it is fixed. Do not fix it by widening the task's file or network access; the scheduled-tasks documentation says to start with the narrowest access that lets the task succeed. A half-migrated task that runs on Monday with a broken prompt is worse than no task. The par logic itself is covered in How Much Food Should You Order? Start With What You Have, and the same review habit applies to the invoice drafts described in How to Use AI to Read Restaurant Invoices Without Losing Control of Costs.
The SOP helper needs a different test: ask it the five questions staff actually ask most, confirm each answer quotes the approved document and page, and confirm it says it cannot answer when the SOP is silent. That approach is described in Gemini 3.8 Live for Restaurant Staff Training; the model changes, the test does not.
Step 4: record the test and set the next review
Write the replacement, the manual test date, the date and result of the first verified scheduled run, and the owner into the register, then set a review date. The changelog and models pages are where OpenAI posts these notices; the models page also records the earlier retirement of gpt-5.4 and gpt-5.4-mini from Codex on August 31, 2026. Recheck those pages on your review date and whenever a task's output changes without an obvious reason.
API-side tools are a different path
The OpenAI API changelog, checked the same morning, has no GPT-5.5 retirement entry, which matches the ChatGPT notice saying the API is unaffected. The practical question for a restaurant is which path each tool uses. A vendor's ordering widget or phone agent that runs on the vendor's API key is on the API path. A scheduled task or Codex script that you signed into with your ChatGPT account is on the ChatGPT path. The same restaurant can have both, and only the ChatGPT path is covered by this notice.
What this post does not cover
It does not cover any custom GPT retirement or timeline, which is documented separately and was not verified for this post. It does not say which model ChatGPT uses by default on your plan, because that is set by OpenAI and your workspace. It does not estimate how long the check takes or what any model costs, and it does not describe the model picker on any particular surface, because those differ. Nothing here was tested against a real restaurant system; the register and tests use invented Bayside Tacos data.
Sources
OpenAI ChatGPT and Codex changelog, entry dated September 14, 2026: learn.chatgpt.com/docs/changelog. Models page, "GPT-5.5 retirement" section and Codex configuration notes: learn.chatgpt.com/docs/models. Scheduled tasks documentation: learn.chatgpt.com/docs/automations. OpenAI API changelog: developers.openai.com/api/docs/changelog. All pages read on September 20, 2026.