Azure OpenAI large models (gpt-5.6-luna/sol, gpt-5.4, o3) hang in OpenCode due to Responses API streaming

[Str0k]

Bug description

When using the native Azure OpenAI provider (azure) in OpenCode, small models like gpt-5-mini work correctly, but larger models (gpt-5.6-luna, gpt-5.6-sol, gpt-5.4, o3) hang indefinitely instead of returning a response.

Steps to reproduce

1. Configure Azure OpenAI provider with a valid resource and API key:

   $env:AZURE_RESOURCE_NAME = "<your-resource-name>"

2. Run OpenCode with a large Azure model:

   opencode run --model azure/gpt-5.6-luna "hola, responde solo OK"

3. The command hangs and never returns (tested up to 5 minutes).

Expected behavior

The model should respond, just like azure/gpt-5-mini does.

Actual behavior

• azure/gpt-5-mini → ✅ works
• azure/gpt-5.6-luna → ❌ hangs forever
• azure/gpt-5.6-sol → ❌ hangs forever
• azure/gpt-5.4 → ❌ hangs forever
• azure/o3 → ❌ hangs forever

Direct API verification

I verified that the Azure OpenAI endpoints work correctly when called directly:

• Responses API (/openai/v1/responses?api-version=v1) with gpt-5.6-luna responds quickly and streams correctly.
• Chat Completions API (/openai/deployments/gpt-5.6-luna/chat/completions?api-version=2025-04-01-preview) also responds correctly.

So the Azure deployment and credentials are fine. The hang only happens through OpenCode.

Root cause hypothesis

OpenCode uses @ai-sdk/azure, which defaults to the Responses API when calling azure(deploymentName). The Responses API streaming path seems to hang for larger models when consumed by OpenCode, even though the same endpoint works fine with direct curl/PowerShell calls.

A likely fix would be to either:

1. Allow users to force Chat Completions API for Azure models (e.g., via provider.azure.options.factory: "chat"), or
2. Make @ai-sdk/azure default to chat completions for models/deployments where Responses API streaming is unreliable.

Environment

• OS: Windows 11
• OpenCode version: 1.17.11 (also tested on earlier versions)
• Azure region: East US 2
• Models deployed: gpt-5-mini, gpt-5.6-luna, gpt-5.6-sol, gpt-5.4, o3

Additional context

I also tried creating a custom provider using @ai-sdk/azure with useDeploymentBasedUrls: true, but it still calls /responses and returns 404 for api-version=2025-04-01-preview because that API version does not support the Responses API on Azure.

The only working workaround right now is to use azure/gpt-5-mini for everything, which is not ideal for complex tasks.

[github-actions]

This issue might be related to an existing open issue. Please check:

• SSE stream silently dropped mid-response hangs session/subagents forever — chunkTimeout has no default on the openai path #37580: Covers a similar "hangs forever" symptom on SSE streams (specifically mentions gpt-5.6-sol). The root cause there is a silently dropped SSE connection with no default chunkTimeout, which may also be a contributing factor here — worth checking if setting provider.azure.options.chunkTimeout in opencode.json changes the behavior for you.

The Responses API vs Chat Completions API routing distinction you identified is a separate and specific concern, so this may still be a distinct bug.

[Str0k]

Update: found the workaround.

Adding useCompletionUrls: true to the Azure provider options forces OpenCode to use the Chat Completions API (sdk.chat(modelID)) instead of the Responses API (sdk.responses(modelID)), and all the previously hanging models now work:

"azure": {
  "options": {
    "useCompletionUrls": true
  }
}

Verified working:

• azure/gpt-5.6-luna ✅
• azure/gpt-5.6-sol ✅
• azure/gpt-5.4 ✅
• azure/o3 ✅

This matches the logic described in #17552. The default sdk.responses(modelID) path hangs for these larger Azure OpenAI models, while sdk.chat(modelID) works correctly.

It would still be helpful if this were better documented or if the Azure provider defaulted to chat completions for models where Responses API streaming is unreliable.

[rekram1-node]

It would still be helpful if this were better documented or if the Azure provider defaulted to chat completions for models where Responses API streaming is unreliable.

Hmm this sounds like an azure side bug or something is misconfigured somewhere. The openai models perform best through responses apis so we don't wanna switch that default.

You say:

I verified that the Azure OpenAI endpoints work correctly when called directly:

How are u verifying? How much context are u sending, i imagine it would be similar in a 10k context scenario

Are u using any plugins?

[Str0k]

@rekram1-node thanks for the follow-up. I did some more controlled testing and I think I can narrow down exactly where the hang comes from.

Verification method

I called the Azure OpenAI endpoints directly with curl/PowerShell against:

• https://rammm-vexium-resource.openai.azure.com/openai/responses?api-version=2025-04-01-preview
• https://rammm-vexium-resource.openai.azure.com/openai/deployments/gpt-5.6-sol/chat/completions?api-version=2025-04-01-preview

Results

Chat Completions API works directly:

POST /openai/deployments/gpt-5.6-sol/chat/completions?api-version=2025-04-01-preview
{
  "model": "gpt-5.6-sol",
  "messages": [{"role": "user", "content": "responde solo OK"}],
  "max_completion_tokens": 100
}

Response: OK

Responses API also works directly, but only if the request body uses input instead of messages:

POST /openai/responses?api-version=2025-04-01-preview
{
  "model": "gpt-5.6-sol",
  "input": [{"role": "user", "content": "responde solo OK"}]
}

Response: OK

However, if I send messages to the Azure Responses API, it returns a 400 Bad Request:

{
  "error": {
    "message": "Unsupported parameter: 'messages'. In the Responses API, this parameter has moved to 'input'.",
    "type": "invalid_request_error"
  }
}

What OpenCode appears to be doing

When useCompletionUrls is not set, OpenCode routes these Azure models through sdk.responses(modelID) but still sends the Chat Completions-shaped payload with messages. Azure rejects that with a 400, and in OpenCode the request seems to hang/never surface the error.

With useCompletionUrls: true, OpenCode uses sdk.chat(modelID), sends messages, and everything works.

Context size

I reproduced the hang with a trivial prompt (3 input tokens, no files, no context), so it is not related to large context.

Plugins

No plugins or MCP servers are configured:

"plugins": [],
"mcpServers": {}

Suggestion

It looks like the Azure provider needs to either:

1. Map messages -> input when calling Azure Responses API, or
2. Default Azure large models to Chat Completions until Azure Responses API accepts the same body shape as OpenAI's Responses API.

Happy to provide more logs or test a patch.