Why does my custom GPT show the old behavior after I edited it?

Check three things: whether you saved the change, whether you published the updated version, and whether you are testing in a new conversation. All three need to be done before the change is visible.

Do I need to republish every time I make a change?

Yes. Each time you update the instructions or knowledge files, you need to save and then publish to make the updated version active for users.

Can I keep the old version of my GPT while testing the new one?

The GPT editor publishes one version at a time. You cannot run the old and new versions in parallel inside ChatGPT. Save the old instructions in an external file before editing so you can restore them if needed.

What if the instructions in the editor look correct but the GPT is still behaving wrong?

Make sure you are in a new conversation, not an old one that carries earlier context. If the problem persists in a new conversation, the issue is in the instructions themselves rather than a publish or save problem.

Custom GPT Not Updating? Check Saved and Published Changes

If your custom GPT changes are not showing up, the usual cause is an unsaved draft, an unpublished version, or a stale conversation. Here is how to check each one.

October 13, 2026Updated October 21, 20264 min read

When you edit a custom GPT and the new behavior does not appear, the change either did not save, did not publish, or is being tested in a conversation that was started before the update. Those are three separate problems with three separate fixes.

Custom GPTs are saved configurations in ChatGPT that you build and maintain in the GPT editor. Changes to the instructions, knowledge files or actions take effect only after you save and publish them. A tool like fenbs keeps a history of your approved instruction versions outside the editor, so you can compare what changed between the version that worked and the one that does not.

Confirm you saved the change

The GPT editor has a Save or Update button. If you made a change and closed the tab or navigated away without saving, the change is gone. Go back into the editor and check whether the current instructions match what you intended. If they do not, re-enter the change and save.

If you are not sure what the last saved version contained, compare the current editor text against any copy you kept outside the editor. This is one reason to save instruction versions in a separate file or tool before editing.

Check that the right version is published

Saving and publishing are different steps. A saved-but-unpublished change is visible to you in the editor but may not be active for users who access the GPT through a shared link. After saving, confirm that you chose the correct audience (just you, or anyone with the link) and completed the publish step.

If you manage multiple GPTs, double-check that you are editing and testing the same one. It is easy to open the wrong GPT, make a change there, and wonder why the one you intended to update is unchanged.

Test in a new conversation

Conversations with a GPT can carry context from earlier in the same session, and that context can mask an instruction change. If you are testing a GPT in the same conversation where you opened it before the update, open a fresh conversation instead. The new conversation starts with no prior context, so the updated instructions are what the model sees.

This is the most common reason an update seems not to work: the tester sends a follow-up in a long conversation rather than starting fresh. A two-second check — open a new chat — eliminates this possibility.

Verify the knowledge files updated

If your GPT uses uploaded knowledge files and you replaced or added a file, check that the file was processed successfully. Large files can take time to process and may show an error if something went wrong. Open the editor and confirm the file is listed and accepted.

If you updated the content of a file but re-uploaded a file with the same name, check that the new file is the active one. Remove the old file explicitly if needed.

How fenbs helps

Keep a record of what changed and when

In fenbs, save each approved version of your GPT instructions as a Skill, with a note on the date and what changed. When a GPT stops working after an update, you have a clear record of what the last working version said. Start with a free account and save your first instruction version today.

Create a free account

Free for one person · No credit card · Works with Claude, ChatGPT, Cursor and more

Clear the browser cache or try a different browser

In some cases, a browser cache issue can serve you a stale version of the ChatGPT interface, making it appear that a GPT has not updated when the change was saved and published correctly. Try opening the GPT in a private browsing window or a different browser. If the update appears there, clearing your browser cache on the main browser should fix it.

This is less common than the other causes but worth a quick check before spending time on the instructions. If the GPT looks different in a fresh browser session than in your usual one, the issue is the cache rather than the setup.

Keep a version outside the editor

The GPT editor does not keep a version history. Once you save a new version, the previous one is gone. If you make a change, find it breaks something, and want to revert, you need to have kept the earlier version somewhere else.

Before editing, copy the current instructions into a separate file or a prompt library. This also makes it easier to compare the current instructions against an earlier version when something unexpected starts happening after an update. Small differences between versions — a changed word, a removed sentence — can have noticeable effects on the output.

How to do this with fenbs

Create a free account and add a project for the work your custom GPT supports.
Before editing the GPT, copy the current instructions into a Skill in fenbs and add a note with today's date.
After making a change, update the Skill with the new version and a short note on what changed and why.
When you test the update, use a new conversation and note the test inputs and outputs in the Skill or project.
If the update causes a problem, open the Skill history in fenbs to find the last version that worked.

For related problems, see custom GPT not following instructions if the GPT has updated but is still behaving incorrectly. For a comparison of platforms, see Claude Projects vs. custom GPTs.