The Model Inventory Sheet – Mathew's Library

01 Why you need the list

On September 14, 2026, OpenAI posted a notice: GPT-5.5 would leave ChatGPT, ChatGPT Work and Codex, OpenAI's coding tool, on every plan from Free to Enterprise. On October 14 it was gone. The API kept it.

When a model retires, your work doesn't stop. It just stops being tested. OpenAI's own notice lists six places a model can sit inside your work:

Workspace defaults. The model everyone in your workspace starts on.

Saved model settings. A model someone picked once and never changed.

Managed configurations. Settings an administrator pushes to everyone's Codex.

Custom agents. Agents your team built that name a model.

Scheduled tasks. Anything that runs by itself on a timer, like a Monday report.

Scripts. Code that still selects a model by name.

And if you never picked a model at all, OpenAI's Codex documentation says you get a recommended one. You were on the default, and the default moves.

02 The sheet, six columns

Make a list. Not of your AI tools. Of your AI workflows. One row for every thing that runs on a model: the Monday report, the custom GPT your sales team uses for proposals, the project with your brand voice in it, the little script someone wrote that calls the API.

Column What to write Example row

What it is The workflow, in plain words The Monday report

Where it lives A chat, a custom GPT, a project, a scheduled task, Codex, or code A scheduled task

Model The model it uses. If nobody picked one, write "default" GPT-5.5

Owner The person who checks it on the day The account lead

Vendor date The vendor's retirement date, if there is one October 14

Last tested The last time someone tested it The day it was set up

Read that row out loud and the problem is obvious. If you can't fill in the model column on a row, that's your first finding.

Where to find the vendor date for code: on the API, the model list has an optional shutdown date field. It holds the date as YYYY-MM-DD, or stays empty when no shutdown date has been announced.

03 An owner for every row

There's one catch on Enterprise or Edu. GPT-6.1 Sol, the replacement, stays off until an administrator turns it on. OpenAI's documentation says changing a default doesn't grant model access, and a model setting in your ChatGPT workspace doesn't carry over to Codex or the API.

So you can switch the default in one place and leave someone else pointing at a model they can't use. That's why every row needs a person who checks it on the day.

04 The five-input test

Before the date, take five real inputs, five weeks of real numbers, and save the outputs you were happy with. Those five are your answer key. When a new model arrives, run the same five through it and put the old and the new side by side. Don't ask which one is smarter. Ask about what you care about:

Is the table there? The format you expect, every time.

Are the numbers the same numbers? Compare them line by line with your answer key.

Is anything added that you didn't give it? A fact, a claim or a promise that wasn't in the input.

Does it still sound like you? Read it as the client would.

If all five pass, update the row: the model and the date. If one fails, fix the instructions and run it again. If it still fails, keep that one workflow on the old model through the API for a while, and write that down too, with its own date.

Test record

Workflow: [NAME]
Old model: [MODEL] New model: [MODEL]
Inputs tested: [5 INPUTS, DATES]
Table there: [YES / NO]
Same numbers: [YES / NO]
Anything added: [NONE / WHAT]
Sounds like us: [YES / NO]
Result: [PASS / FIXED INSTRUCTIONS / KEPT ON API UNTIL DATE]
Row updated on: [DATE] by [OWNER]

Sources

OpenAI: ChatGPT & Codex changelog OpenAI: Codex models OpenAI: Workspace model availability OpenAI API: Deprecations

Mathew's LibraryLibrary card

Free access

Keep reading free with your library card

One email opens this template and all 86 resources in Mathew's library. No account, no password.