### What happened?

DeepSeek V4 Pro and V4 Flash (released 2026-04-24) support **graded reasoning effort** via the `reasoning_effort` parameter — both `"high"` and `"max"` (with `xhigh` → `max`, `low`/`medium` → `high` for compatibility). This is officially documented at: 👉 https://api-docs.deepseek.com/guides/thinking_mode#thinking-mode-toggle-and-effort-control

However, LiteLLM's `DeepSeekChatConfig.map_openai_params()` in `litellm/llms/deepseek/chat/transformation.py` **discards the value** and always maps to `thinking: {"type": "enabled"}`, regardless of the requested effort level:

```python elif reasoning_effort is not None and reasoning_effort != "none": optional_params["thinking"] = {"type": "enabled"} # value thrown away! ```

This means: - `reasoning_effort="max"` → `thinking: {"type": "enabled"}` (effort lost) - `reasoning_effort="high"` → `thinking: {"type": "enabled"}` (effort lost) - User has no way to request `max` reasoning via LiteLLM

### Secondary issue: OpenRouter path does not apply DeepSeekChatConfig

When using `openrouter/deepseek/deepseek-v4-pro`, the provider is `openrouter` and `OpenrouterConfig` is used instead of `DeepSeekChatConfig`. This means: …