Some models such as 5.6-terra, 5.6-sol, 6.1-sol, 6.1-terra fail as Model Provider

Hello,

I'm experimenting with the assistant in Reality Composer Pro 3.1.
On the OpenAI dashboard I added an API key to a project. I was able to add my OpenAI as a Model Provider using the API key and configure RCP model list.

Initially I used gpt-5.5 and it worked. I was able to generate some assets: 3D models, PBR materials, ShaderGraph shaders, and textures.

gpt-5.5 runs into a high cost so I was suggested to use 6.1-terra or 6.1-sol, which are more cost effective. terra and sol do not work, the query is failed early and all I get is an error in the prompt. The feedback file generated contains this:

{
  "modelProvider" : "openai",
  "modelName" : "gpt-6.1-terra (openai)",
  "prompt" : "Create a 3d model of an art deco desk lamp, with textures. Green body, smokey glass shade.\n"
}

"gpt-6.1-terra" does match the actual name of the model.

In RCP3 I set-up the models like this:

The error appears quite fast, in just a few seconds.

I aso tested 5.6-terra or 5.6-sol, same result.

In the OpenAI dashboard an 400 error is logged:

400 - Invalid request
Nov 16, 2026, 08:09 UTC
rcp3-model-provider

Cause: The request did not meet the endpoint's requirements.

Solution: Check the request parameters against the API reference

(rcp3-model-provider is the OpenAI project name).
In Usage I was able to see some of the queries, while others are not appearing (we are not sure the reason why some show up while others don't).

I would like to debug this issue but ot sure where to start. Some hypothesis:

  • the request is malformed or not compatible with terra/sol models.
  • terra/sol models lack some feature that gpt-5.5 has, but which one I don't know.

There is not much information on query formatting and recommended models.

Any help is appreaciated.

This other thread is somewhat related but is not the same problem. It is about models not being listed in the UI, but it seems this issue was resolved at some point. I can select the models but they just do not work.
https://developer.apple.com/forums/thread/839809?answerId=899528022

Comparing the responses from OpenAI endpoint in Postman, I oberved that there is a significant difference.
My test case uses https://api.openai.com/v1/chat/completions and a request to check if a tool calling works or not. The check in itself is simple on purpose.

  • gpt-5.5: "reasoning_effort": "none" is not required, response is 200 OK, and response's body confirms the tool was called successfully.
  • gpt-6.1-terra: "reasoning_effort": "none" is required, otherwise you get 400 Bad Request.

the 400 error is:

{
    "error": {
        "message": "Function tools with reasoning_effort are not supported for gpt-6.1-terra in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'.",
        "type": "invalid_request_error",
        "param": "reasoning_effort",
        "code": null
    }
}

An error is also log for gpt-6.1-terra in OpenAI dashboard.
I suspect that RCP3 is calling a tool but this setting is not present in the request. 5.5 does not need but terra/sol models do require it and because it is absent it fails.

There is no way to add extra parameters to the request from within RCP3 UI. In theory it should be possible to point to a local proxy and reqrite the request and then pass it back to RCP, but this setup seems more complex and it may fail (if headers get altered in an incorrect way, or streaming is required).

I would appreciate if Apple engineers looked into this.