Custom GPT Actions not triggered on GPT-5.5 Instant, but work on GPT-5.3 Instant and GPT-5.5 Thinking

ChatGPT Bugs
, , , ,
1

Hi,

I’m seeing what looks like a Custom GPT Actions regression affecting newer Instant models.

I have a Custom GPT that must call a database action before answering any stock or security analysis request. The action is configured through an OpenAPI schema and points to my backend.

The same Custom GPT, same instructions, same OpenAPI schema, same backend, same authentication, and same prompt behave differently depending on the selected model.

Observed behavior:

  • GPT-5.3 Instant: works correctly. The GPT calls the action before answering.
  • GPT-5.4 Instant: does not call the action. It returns the configured fallback response.
  • GPT-5.5 Instant: does not call the action. It returns the configured fallback response.
  • GPT-5.5 Thinking: works correctly. The GPT calls the action before answering.

Example prompt:

Analyze MSFT

Expected behavior:

The GPT should call the Custom GPT Action before producing any visible answer.

Actual behavior on GPT-5.4 Instant and GPT-5.5 Instant:

The model does not call the HTTP endpoint and returns the configured fallback response.

Actual behavior on GPT-5.3 Instant and GPT-5.5 Thinking:

The GPT Builder Preview debug panel shows that the HTTP endpoint is called and a response is received.

The backend is reachable and working, because the exact same action is called successfully from GPT-5.3 Instant and GPT-5.5 Thinking.

I have already hardened both the GPT instructions and the OpenAPI schema to make the action requirement explicit.

The operation summary is:

MANDATORY first DB fetch for any security analysis

The operation description starts with:

ALWAYS call this action before answering any user request about a stock, ETF, bond, REIT, security, ticker, company, or security number. Do not answer from memory. Do not produce visible text before this action.

The OpenAPI JSON validates successfully, the action imports correctly, and the backend works when called by other models.

This appears to be a Custom GPT Actions tool-selection regression in GPT-5.4/5.5 Instant, not an API, backend, authentication, or OpenAPI schema issue.

Based on my tests, the behavior seems to have changed after GPT-5.3 Instant. GPT-5.3 Instant still calls the action correctly, while GPT-5.4 Instant and GPT-5.5 Instant do not.

Questions:

  1. Is there a known change in GPT-5.4/5.5 Instant tool/action selection behavior for Custom GPT Actions?
  2. Is there a recommended way to force or strongly bias an action call in newer Instant models when the action is mandatory for the task?
  3. Are there any known differences between GPT-5.3 Instant and GPT-5.4/5.5 Instant regarding Custom GPT Actions?

I can provide redacted screenshots from GPT Builder Preview showing:

  • GPT-5.3 Instant calling the endpoint successfully
  • GPT-5.4 Instant not calling the endpoint
  • GPT-5.5 Thinking calling the endpoint successfully
2

Additional update:

I found another related issue with the Custom GPT recommended-model setting.

This is separate from the GPT-5.4/5.5 Instant Action-calling regression itself.

In the GPT Builder configuration, I selected GPT-5.3 Instant as the recommended model.

The GPT page displays:

“Using the creator’s recommended model: GPT-5.3 Instant”

However, inside the actual GPT chat, the model selector can still show GPT-5.4 as the active model.

I also noticed this inconsistency inside GPT Builder Preview itself, not only in the published GPT. The Preview area displays that it is using the creator’s recommended model, GPT-5.3 Instant, but the model selector in the same Preview session can show GPT-5.4 as the active model.

If I manually switch the active model inside the chat from GPT-5.4 to GPT-5.3, the Custom GPT Action starts working correctly again.

So there appear to be two separate but related problems:

  1. GPT-5.4/5.5 Instant do not reliably trigger the Custom GPT Action.
  2. The Custom GPT does not appear to reliably start on the configured recommended model, or the UI/model routing label is inconsistent.

The second issue does not explain why GPT-5.4/5.5 Instant fail to call Actions. Rather, it prevents GPT-5.3 Instant from being a reliable workaround, because the GPT may still open on GPT-5.4 unless the user manually switches models.

Current observed recommended-model behavior:

  • GPT Builder recommended model: GPT-5.3 Instant
  • GPT page banner: “Using the creator’s recommended model: GPT-5.3 Instant”
  • Actual model selector in chat: can show GPT-5.4
  • Builder Preview: can also show the same inconsistency
  • Manual switch to GPT-5.3: Custom GPT Action works again

This makes the issue harder to work around, because I cannot simply set the Custom GPT recommended model to GPT-5.3 Instant and rely on it being used automatically.