AI Chat with GitHub Copilot fails with: 400 The requested model is not supported

[steffentemplin]

Bug Description:

Thia AI fails on every attempt to start a chat with error: "400 The requested model is not supported."

Steps to Reproduce:

1. Install and launch Thia IDE
2. Enable AI
3. Authenticate GH Copilot
4. Configure Models and Aliases
5. Open AI Chat and start a conversation with @Universal

Additional Information

• Operating System: Windows 11
• Theia Version: 1.68.201

2026-02-25T20:43:53.474Z root ERROR Error in OpenAI chat completion stream: Error: 400 The requested model is not supported.
    at o.generate (C:\Users\user\AppData\Local\Programs\TheiaIDE\resources\app.asar\lib\backend\845.js:8:6764554)
    at J.makeStatusError (C:\Users\user\AppData\Local\Programs\TheiaIDE\resources\app.asar\lib\backend\845.js:8:6753870)
    at J.makeRequest (C:\Users\user\AppData\Local\Programs\TheiaIDE\resources\app.asar\lib\backend\845.js:8:6757709)
    at process.processTicksAndRejections (node:internal/process/task_queues:105:5)
    at async E._createChatCompletion (C:\Users\user\AppData\Local\Programs\TheiaIDE\resources\app.asar\lib\backend\845.js:8:6825434)
    at async E._runChatCompletion (C:\Users\user\AppData\Local\Programs\TheiaIDE\resources\app.asar\lib\backend\845.js:8:6808266)
2026-02-25T20:43:53.482Z root ERROR Error handling chat interaction: Error: 400 The requested model is not supported.
    at o.generate (C:\Users\user\AppData\Local\Programs\TheiaIDE\resources\app.asar\lib\backend\845.js:8:6764554)
    at J.makeStatusError (C:\Users\user\AppData\Local\Programs\TheiaIDE\resources\app.asar\lib\backend\845.js:8:6753870)
    at J.makeRequest (C:\Users\user\AppData\Local\Programs\TheiaIDE\resources\app.asar\lib\backend\845.js:8:6757709)
    at process.processTicksAndRejections (node:internal/process/task_queues:105:5)
    at async E._createChatCompletion (C:\Users\user\AppData\Local\Programs\TheiaIDE\resources\app.asar\lib\backend\845.js:8:6825434)
    at async E._runChatCompletion (C:\Users\user\AppData\Local\Programs\TheiaIDE\resources\app.asar\lib\backend\845.js:8:6808266)

settings.json:

{
    "ai-features.AiEnable.enableAI": true,
    "ai-features.chat.bypassModelRequirement": true,
    "ai-features.chat.defaultChatAgent": "Universal",
    "ai-features.copilot.models": [
        "claude-haiku-4.5",
        "claude-sonnet-4.5",
        "claude-opus-4.5",
        "gemini-2.5-pro",
        "gpt-4.1",
        "gpt-4o",
        "gpt-5-mini",
        "gpt-5.2"
    ],
    "ai-features.agentSettings": {
        "Chat Session Naming": {
            "languageModelRequirements": [
                {
                    "purpose": "chat-session-naming",
                    "identifier": "default/summarize"
                }
            ]
        },
        "Code Completion": {
            "languageModelRequirements": [
                {
                    "purpose": "code-completion",
                    "identifier": "default/code-completion"
                }
            ]
        },
        "Architect": {
            "languageModelRequirements": [
                {
                    "purpose": "chat",
                    "identifier": "default/code"
                }
            ]
        },
        "Universal": {
            "languageModelRequirements": [
                {
                    "purpose": "chat",
                    "identifier": "copilot/claude-sonnet-4.5"
                }
            ]
        }
    },
    "ai-features.languageModelAliases": {
        "default/code": {
            "selectedModel": "copilot/claude-sonnet-4.5"
        },
        "default/universal": {
            "selectedModel": "copilot/claude-opus-4.5"
        },
        "default/code-completion": {
            "selectedModel": "copilot/claude-sonnet-4.5"
        },
        "default/summarize": {
            "selectedModel": "copilot/claude-opus-4.5"
        }
    }
}

[ndoschek]

Hi @steffentemplin, thanks for reporting this!
The issue is that the default model list we ship is based on what GitHub Copilot advertises as available. However, not all of these models are currently supported anymore through the Copilot API for third-party clients it seems. This used to work, so we need to investigate what changed.

As a workaround, the following models work for me today: gpt-4o and gpt-4o-mini.
You can adjust the model list in your settings under ai-features.copilot.models.

We’re starting to look into this. Sorry for the inconvenience in the meantime!

Edit: We identified the root cause: the currently hardcoded model list included models that aren't available on all Copilot plans (see supported models per plan). We're working on a fix that dynamically discovers available models from the Copilot API based on your subscription, so only models you can actually use will be shown.

[fipro78]

@ndoschek
Just to add some information. I have a subscription with access to additional models. They can be accessed at least in VS Code, GitHub and Copilot CLI. But in Theia only gpt-4o and gpt-4o-mini work as you already mentioned. I even tried locally to update the openai version in the ai-copilot and the ai-openai package to the most current 6.29.0 and tried to adapt the usage of the Response API for the other models. But the error message stays the same.

Is there maybe a limitation which models are accessible via openai API in Copilot? I haven't found any information about that.

[sgraband]

@ndoschek can we raise the priority on this? At the very least we should remove models that do not work from the default models list or show them as unready.

Maybe this has something to do with our Github App, where we login the Users, can we maybe specify models there? I do not know who has access to that though.

[ndoschek]

@fipro78 Thanks for the additional info and for trying the openai upgrade,that's helpful to know.

@sgraband Yes, this is already planned for this month's release.
I don't think we specify anything there, but I'll rediscuss this with Philip tomorrow and get back to you.

[ndoschek]

Turns out it most likely has to do with how our GH app is configured. I am looking into this further.

[ndoschek]

I found the root cause: we were using an authorized GitHub App for the Copilot device flow, but the Copilot API now only accepts tokens from OAuth Apps, this is how all working third-party Copilot integrations (like for example OpenCode) authenticate. GitHub App tokens are simply not recognized by the Copilot backend, while OAuth App tokens are accepted directly as Bearer tokens. We registered a new OAuth App today and my first tests work. I am wrapping up the PR and plan to open it tomorrow 👍

[ndoschek]

Hi again everyone, we just merged a PR that introduces automatic model discovery and switches the Copilot integration to an OAuth App (for details please see eclipse-theia/theia#17198).
This will be available with the upcoming release 1.70.x (scheduled for this Thursday).

What changed
Models are now discovered dynamically via GET /models instead of being hardcoded. We also migrated from the previous GitHub App (Iv-prefixed) to a new OAuth App (Ov-prefixed, Theia Copilot OAuth Access). Existing tokens will require a one-time re-authentication.

Current limitation
The OAuth App hasn’t been fully whitelisted by GitHub for Copilot Chat model access yet. As a result, auto-discovery currently returns only the base models (gpt-4o, gpt-4o-mini, gpt-3.5-turbo). We've applied to the Copilot Partner Program; once approved, additional subscription models (like Claude, Gemini, GPT-5, etc.) will become available automatically without further changes.

Related discussions

• https://github.com/orgs/community/discussions/187602
• https://github.com/orgs/community/discussions/187449

We will also update the documentation (in theia-website) accordingly.