## Check for existing issues

- [x] I searched the existing issues and did not find a duplicate covering both the model-scoped `cachedContents` URL and malformed authentication header.

Related:

- #23846 - #23928

## What happened?

Gemini explicit context caching fails when the SDK uses a custom `api_base` implementing the standard Gemini REST API.

When a prompt contains an Anthropic-style `cache_control` marker and exceeds Gemini's minimum cache size, LiteLLM makes a secondary request to the named-cache API. For a custom Gemini base URL, LiteLLM currently:

1. sends the request to: ```text {api_base}/models/{model}:cachedContents ``` 2. converts the Gemini API-key header dictionary into a malformed Bearer value: ```text Authorization: Bearer {'x-goog-api-key': '<redacted>'} ```

The standard Gemini named-cache API uses the collection endpoint:

```text {api_base}/cachedContents ```

The model belongs in the create request body, for example:

```json {"model": "models/gemini-2.5-flash"} ```

This affects standards-compatible Gemini proxies such as Bifrost. The normal `generateContent` request succeeds through the same custom `api_base`; only the explicit context-…