AI Provider (Custom / OpenAI-compatible): temperature parameter causes errors with GPT-5+ models

Technical assistance
1

Infos:

  • Used Zammad version: 7.1.3-1789383560.f56648e.el9
  • Used Zammad installation type: package
  • Operating system: Ubuntu
  • Browser + version: Firefox

Description:
When using the AI Provider feature with the Custom (OpenAI-compatible) option pointing to Azure AI Foundry, requests fail for OpenAI models from GPT-5 onward (including reasoning models), because Zammad always sends a temperature value in the request body.

Expected behavior:

  • Zammad should allow the request to succeed with these models.

Actual behavior:

Request sent by Zammad

{
  "model": "gpt-5.6-luna",
  "messages": [
    { "role": "user", "content": "Hello" }
  ],
  "temperature": 0.1,
  "stream": false
}

Response from the API (HTTP 400)

{
  "error": {
    "message": "Unsupported value: 'temperature' does not support 0.1 with this model. Only the default (1) value is supported.",
    "type": "invalid_request_error",
    "param": "temperature",
    "code": "unsupported_value"
  }
}

Problem

Starting with GPT-5, OpenAI models no longer accept custom values for temperature (only the default value of 1 is supported). Since Zammad hard-codes temperature: 0.1 in the request, every call to these models fails, making it impossible to use the AI features with current OpenAI models through the custom provider.

Steps to reproduce the behavior:

  • Create a gpt-5+ model deplyment in Microsoft AI Foundry

  • Create a custom AI provider in Zammad, using the Foundry URL (http://******.openai.azure.com), Model and Token

  • Save the changes.

2

Normally, we have already added a handling for that in the different providers, but maybe some edge cases are not working. We will check it.

3

Is it failing directly on save or later when you want to use a feature? Did you change the configuration from another value?

4

Or which specific model are you using?

5

You should maybe also try first 7.2 because it could be that we already improved the situation:

6

it is failing immediately after I save the provider settings

7

Is that also for the latest 7.2.2? The productionlog might contain useful information.

8

but this commit is for anthropic models. I’m having this problem with all the GPT models after GPT-5.

9

no, my version is 7.1.3

10

You might want to test a supported stable. Or areyou certain nothing changed? Apart of CVEs of course.

11

The request with "content": "Hello" and temperature: 0.1 is not a real AI request. It is a test Zammad sends when you save the provider settings, to find out whether the model accepts a temperature. A 400 with unsupported_value for temperature is the expected answer: Zammad then remembers that the model does not support it and leaves the temperature out of all later requests. That test request is logged because it failed, which is why it shows up right after saving.

So two questions:

  1. When you click save, do you get an error message in the UI, and do the settings stay saved?
  2. Does an actual AI feature (e.g. ticket summary) fail afterwards? If so, please share that log entry, it will contain the real prompt, not “Hello”.

The commit I linked is titled Anthropic, but the change is in code shared by all providers including Custom OpenAI. On 7.1 it matters for GPT models after GPT-5, so please try 7.2.2.

12

ok. I understood. Thank you for your help