Measuring Claude 4.7's Tokenizer Cost

Claude Opus 4.7's new tokenizer uses 1.0–1.35x more tokens according to Anthropic's documentation. Actual measurements confirmed cases exceeding that upper bound in technical documents and real-world Claude Code content.

Measured Results

• Average across 7 real-world Claude Code content samples: 1.325x
• Actual CLAUDE.md file: 1.445x
• Technical documentation (English): 1.47x
• Code and English-heavy content increased up to 1.20–1.47x
• In contrast, CJK (Japanese, Chinese) showed almost no change, staying around 1.01x

What Changed

• Tokens tended to be split more finely for English and code patterns.
• Real-world inputs like code, Markdown, Git logs, and stack traces were particularly affected.
• However, token counts alone don't reveal which slots in the vocabulary actually changed.

Instruction Following Measurement

• On a 20-question IFEval sample, the strict criterion improved by +5pp: 4.6: 85% → 4.7: 90%
• The loose criterion showed almost no change
• The improvement is small but consistent in direction, with 4.7 being especially more stable on precise formatting requirements
• However, with N=20, the effect size remains uncertain

Impact on Claude Code Session Costs

• Assuming an 80-turn long session, 4.7 could increase per-session cost by roughly 20–30% compared to 4.6
• This is because, even with the same pricing, more input tokens lead to more cache reads and cache writes
• For Max plan users, the 5-hour rate limit could also be consumed proportionally faster

Conclusion

• 4.7 raised token costs, but gained a small yet real improvement in strict instruction following.
• Whether that improvement justifies 1.3–1.45x more tokens depends on the type of work.
• For English- and code-heavy Claude Code workflows, it's safer to budget costs based on the upper end of the range.