## Summary

When LiteLLM is used as a **Bedrock passthrough** proxy (`/bedrock/.../invoke-with-response-stream`, e.g. Claude Code with `CLAUDE_CODE_USE_BEDROCK=1`), Anthropic cache-creation tokens written with a **1-hour TTL** (`cache_control: {"type":"ephemeral","ttl":"1h"}`) are billed at the **5-minute** rate.

Bedrock **does** return the split in `usage.cache_creation` (`ephemeral_5m_input_tokens` / `ephemeral_1h_input_tokens`), but the passthrough path keeps only the aggregate `cache_creation_input_tokens` and applies `cache_creation_input_token_cost` (5m) to all of it. `cache_creation_input_token_cost_above_1hr` is never applied.

> The standard chat path (`generic_cost_per_token` → `_calculate_cache_creation_cost`) already reads `cache_creation_input_token_cost_above_1hr` (covered by `tests/test_litellm/test_bedrock_anthropic_1hr_cache_pricing.py`). **The gap is specific to the Bedrock passthrough route.**

## Environment

- LiteLLM **v1.85.0**, Bedrock passthrough endpoint - Model: **Claude Opus 4.8** on Bedrock (Application Inference Profile) - Cost map is correct: `global.anthropic.claude-opus-4-8` → `cache_creation_input_token_cost=6.25e-6`, `cache_creation_input_token_c…