## Problem

`AnthropicConfig` routes `response_format` requests by checking whether the model name contains a hardcoded substring set. When `claude-opus-4-8` was added to the model-cost map with `supports_output_config: true`, the hardcoded routing set was never updated, so every request to `claude-opus-4-8` with `response_format` falls into the tool-call emulation path instead of the native `output_format` API.

Two additional gaps: `eager_input_streaming` is silently dropped from tool definitions because it is absent from both the `_map_tool_helper` passthrough blocks and the `AnthropicMessagesTool` TypedDict. Citation-enriched responses omit character offset fields (`supported_text_start_char_index` / `supported_text_end_char_index`) that downstream consumers need to locate cited text.

## Proposed fix

Replace the hardcoded substring set with a call to `_supports_model_capability(model, "supports_output_config")` (already exists, already correct). Add `eager_input_streaming` passthrough in `_map_tool_helper` and to `AnthropicMessagesTool`. Compute and attach char offsets in `extract_response_content`.

## Files

- `litellm/llms/anthropic/chat/transformation.py` - `litellm/types…