Claude Opus 4.7 tokenizer hikes prompt costs

This looks less like a tokenizer tweak and more like a hidden price increase for anyone running long, cache-heavy Claude workflows.

• –Real-world code-heavy inputs like `CLAUDE.md` and technical docs landed near the top of Anthropic's stated range, so developers should plan for the worst case, not the midpoint.
• –Prompt caching still works, but larger prefixes mean more cache-write and cache-read tokens every turn, which shortens Max windows and pushes rate limits sooner.
• –The measured upside is narrow: about +5 percentage points on a small IFEval sample, enough to matter, not enough to call it a clean trade.
• –CJK content barely changed, so the cost hit is concentrated on English/code-heavy users rather than global text workloads.
• –Bottom line: if your Claude usage is mostly coding, this is a material cost and throughput regression unless the extra literalness saves enough human time to justify it.