I Measured Claude 4.7's New Tokenizer. Here's What It Costs You.

Claude Opus 4.7 ships with a new tokenizer that Anthropic claims uses 1.0–1.35x more tokens than 4.6. Empirical measurements on real Claude Code content (CLAUDE.md files, code diffs, stack traces, prompts) show the actual ratio is 1.325–1.47x, consistently near or above the top of Anthropic's stated range. English and code content is hit hardest (1.29–1.47x), while CJK content barely changes (1.01x). A detailed cost breakdown for an 80-turn Claude Code session shows roughly 20–30% higher per-session cost due to the tokenizer alone, with prompt cache reads being the dominant driver. A small IFEval benchmark test (N=20) shows a +5pp improvement in strict instruction-following for 4.7 over 4.6, though the sample size is too small to be conclusive. The practical takeaway: plan around the upper end of Anthropic's documented range, not the middle, and expect billing step-changes when switching model versions.