#33075 [Bug] Claude Code auto classifier fails for GPT 5.6 models: Anthropic /v1/messages does not translate stop_sequences -> stop for OpenAI-format providers (400 "Unknown parameter: 'stop_sequences'")
## What happened?
Calling LiteLLM's Anthropic-compatible `/v1/messages` endpoint with `stop_sequences`, when the target deployment is an **OpenAI-format provider** (e.g. `azure_ai` / Azure AI Foundry), returns **HTTP 400**. LiteLLM forwards the Anthropic field name `stop_sequences` **verbatim** to the upstream instead of translating it to the OpenAI `stop` parameter, and the upstream rejects the unknown field:
```json {"error":{"message":"litellm.BadRequestError: AzureException BadRequestError - Unknown parameter: 'stop_sequences'.. Received Model Group=<model>\nAvailable Model Group Fallbacks=None","type":"invalid_request_error","param":"stop_sequences","code":"400"}} ```
The same request **succeeds** when `stop_sequences` is omitted, and **succeeds** when the backing deployment is an Anthropic-native provider (Bedrock/Anthropic), where no translation is needed. The OpenAI Chat Completions path already accepts the native `stop` field for the same deployment, so the backend itself has no issue with stop sequences ā the defect is a missing `stop_sequences` ā `stop` mapping in the Anthropic-`/v1/messages` ā OpenAI-provider adapter.
**Impact:** any Anthropic-API client that sets sā¦