## Background

litellm ships two model-pricing JSONs: - Root `model_prices_and_context_window.json` (39k+ lines, source of truth maintained by reviewers + automated bumps) - `litellm/model_prices_and_context_window_backup.json` (mirror used when no remote map is available — i.e. CI default and when `LITELLM_LOCAL_MODEL_COST_MAP=True`)

The backup is supposed to be a periodic snapshot of the root. In practice it drifts: features get merged into the root file (e.g. `supports_prompt_caching: true` on the entire xAI grok-3 family) without the backup being resynced. CI tests that read via `litellm.get_model_cost_map(url=\"\")` see the stale backup; locally, contributors with `LITELLM_LOCAL_MODEL_COST_MAP=True` see the same; production deployments pull the root from main and don't notice.

## Concrete recent example (PR #27074)

`xai/grok-3-mini-beta` had `supports_prompt_caching: true` in the root JSON but **not** in the backup. Effects: 1. `test_xai.TestXAIChat::test_prompt_caching` skipped on `main` for months (`supports_prompt_caching(\"xai/grok-3-mini-beta\", None)` returned `False` from the stale backup). 2. PR #27074's commit [`108b87fb24`](https://github.com/BerriAI/litellm/commi…