MiniMax's OpenAI-compatible endpoint has a known issue where tool calls
are sometimes returned as XML content (<minimax:tool_call>) instead of
structured tool_calls. Their Anthropic Messages API endpoint handles
this correctly.
Changes:
- Switch MiniMax/MiniMax-CN base URLs to Anthropic endpoints
(api.minimax.io/anthropic, api.minimaxi.com/anthropic)
- Return api_mode='anthropic_messages' for MiniMax providers
- Add third_party flag to build_anthropic_client() to skip OAuth
detection, Anthropic beta headers, and Claude Code user-agent
- Guard normalize_model_name() to not mangle non-Claude model names
(MiniMax-M2.5 would incorrectly become MiniMax-M2-5)
- Skip Anthropic credential refresh for third-party providers
- Show provider-appropriate 401 diagnostics for third-party providers
- Add missing MiniMax models (M2.1-highspeed, M2) to metadata/catalog
- Update tests for new URLs and third_party kwarg
MiniMax's Anthropic compat layer supports:
- Tool calling (text, tool_use, tool_result)
- Thinking/reasoning (interleaved thinking)
- Explicit prompt caching (cache_control with 5min TTL)
- Streaming
- 204,800 token context for all M2/M2.1/M2.5 models
Ref: https://platform.minimax.io/docs/api-reference/text-anthropic-api
teknium1
3ea53b8eed