Two related problems in `model_prices_and_context_window.json`, found while building a cache-cost analyzer that reads this file as a source of truth. The first is a wrong value, the second is a coverage gap. Both surface through `get_prompt_cache_min_tokens()`.

## 1. `claude-fable-5` contradicts itself

| key | `prompt_cache_min_tokens` | |---|---| | `claude-fable-5` | **512** | | `anthropic.claude-fable-5` | 1024 | | `us.anthropic.claude-fable-5` | 1024 | | `eu.anthropic.claude-fable-5` | 1024 | | `global.anthropic.claude-fable-5` | 1024 |

Anthropic publishes 512 for Fable 5, which the direct entry already has. The four platform entries say 1024.

Effect: on those four, a 700-token prefix is treated as too short to cache. `is_prompt_caching_valid_prompt()` returns `False`, the caller skips caching, and a prefix that would have cached is re-sent at full rate on every request.

I checked every other Anthropic model in the file for the same kind of internal disagreement. This is the only one. Every other recorded minimum is correct, including the non-monotonic ones (512 on Opus 5, 1024 on Opus 4.8 and Sonnet 5, 2048 on Opus 4.7 and Haiku 3.5, 4096 on Opus 4.6, Opus 4.5 and Haiku 4.…