## Summary

LiteLLM proxy's Anthropic-compatible `/v1/messages` endpoint does **not** trigger the `async_pre_call_hook` of registered `CustomLogger` callbacks. The same callback registered the same way **does** trigger correctly when calling `/v1/chat/completions` (OpenAI endpoint).

This makes it impossible to use callbacks to modify the `model` parameter for requests using the Anthropic protocol — for example, content-based model routing in a multi-model proxy setup.

## Reproduce

### 1. Minimal callback (`router_callback.py`)

```python import logging from litellm.integrations.custom_logger import CustomLogger

logger = logging.getLogger("repro")

class RouterCallback(CustomLogger): async def async_pre_call_hook(self, user_api_key_dict, cache, data, call_type): logger.warning("pre_call_hook fired: model=%s call_type=%s", data.get("model"), call_type) if data.get("model") == "qwen3.5-35b": data["model"] = "devstral-small-2" logger.warning("model swapped -> %s", data["model"]) return data

router_instance = RouterCallback() ```

### 2. LiteLLM config (`litellm_config.yaml`)

```yaml model_list: - model_n…