Auth error with GPT 5.6-xxx models

[kvokka]

Description

With the GPT 5.6 model, the request does not go through, while with the GPT 5.5, everything works correctly;

opencode.json (cut)

{
...
    "openai": {
      "models": {
        "gpt-5.6-luna": {
          "limit": {
            "context": 258000,
            "output": 128000
          },
          "name": "GPT 5.6 Luna",
          "options": {
            "reasoningEffort": "high",
            "textVerbosity": "medium"
          },
          "variants": {
            "medium": {
              "reasoningEffort": "medium"
            },
            "low": {
              "reasoningEffort": "low"
            },
            "high": {
              "reasoningEffort": "high"
            },
            "xhigh": {
              "reasoningEffort": "xhigh"
            }
          },
          "reasoning": true
     },
      "name": "OpenAI",
      "npm": "@ai-sdk/openai",
      "options": {
        "apiKey": "{env:MY_CLI_LLM_PROXY_API_KEY}",
        "baseURL": "{env:MY_CLI_LLM_PROXY_URL}"
      },
}

The older models work correctly (as well as custom proxy url, since codex cli is working fine with it), but with Opencode, I face a tricky auth err.

Models in the config were added just to avoid any ambiguity

Plugins

none

OpenCode version

1.17.15

Steps to reproduce

1. Pick up the GPT5.6 model under the ChatGPT subscription Auth
2. Say hi

Screenshot and/or share link

No response

Operating System

Ubuntu 24

Terminal

iTerm2

[github-actions]

This issue might be a duplicate of or related to an existing issue. Please check:

• Custom @ai-sdk/google provider loads correctly but drops API key at runtime when using custom baseURL #21738: Custom provider drops API key at runtime when using custom baseURL — the auth error symptom when using a custom proxy URL is similar to what you're experiencing, though that report is for @ai-sdk/google. If the same root cause (apiKey not propagated through bundled provider path) applies to @ai-sdk/openai, it may explain why GPT 5.5 works but GPT 5.6 does not (if the two models use different underlying request paths).

[Dayclone]

I get something similar with my GPT subscription but I suspect its not GA yet. Bad Request: {"detail":"The 'gpt-5.6' model is not supported when using Codex with a ChatGPT account."}

[kvokka]

But codex over CliProxy works, meaning that the interface is the same (it didn't change, just checked).
And GPT 5.5 works with the same setup. Only GPT 5.6-xxx are not

[emcd]

The difference in the error response could be the Opencode version. With 1.17.18, I see the same response as @Dayclone :

Bad Request: {"detail":"The 'gpt-5.6-sol' model is not supported when using Codex with a ChatGPT account."}

The original report was against an older version (1.17.15), which might not have the same logic, especially if everything needed about the new models (e.g., Oauth connectivity) does not live on Models.dev.

[rekram1-node]

there is an issue specifically w/ luna it seems we will track it here:
#36140