Mcp tools not being invoked reliably since GPT-5.1

ChatGPT Apps SDK
1

Hi

I have a ChatGPT app (using app sdk) that utilises a number of mcp tools, some of which chain together.

since moving to ChatGPT 5.1 the reliability of these tools have degraded.

examples

One tool provides product results for multiple searches, previously the model could select the best products for the recipe from these, but now it just shows a list of all the results. My tool description informs this should be the behaviour)

Previously I had two Tools chained, one to find the products which then directed the llm to reason the best products before passing on the selected products to a second tool that used a template resource to render them in our custom ux.

this now works unreliable despite the prompts being the same. Intermittently the first tool will complete without the second tool being invoked.

all of this was stable pre 5.1.

Is this a known issue? Can anyone provide guidance?

Thanks

4

Unfortunately, I’m encountering similar reliability scenarios. Especially when it comes to hallucinating tool names. We have an overview section for the tool suite, with clear instructions that only the tools listed are available, but then it tries to guess their names and persists 4-6 times with different variants.

I’m trying to use tactics from here, but can’t say I figured it out yet