Claude 4.7 Tokenizer Cost Measurement Results

Claude Opus 4.7 generates 1.3–1.45x more tokens on average than 4.6, increasing per-session costs by 20–30% even under the same pricing structure.

The measurement was conducted using the Anthropic API's POST /v1/messages/count_tokens. 7 real Claude Code usage samples were compared against 12 artificial samples covering English, code, structured data, CJK, emoji, and math symbols.

• Weighted average for real-usage samples: 1.325x increase (8,254 → 10,937 tokens)
• Average for English/code content: 1.345x
• CJK (Chinese, Japanese, Korean): 1.01x, almost no change
• Technical documentation: 1.47x, Shell script: 1.39x, TypeScript: 1.36x
• English prose: 1.20x, JSON: 1.13x

The increase pattern was especially pronounced in English and code. Repeated strings, keywords, identifiers, and the like were broken down more finely, causing the same text to be represented with more tokens. In contrast, CJK, emoji, and symbols showed almost no change.

Instruction compliance showed a modest improvement in IFEval sample tests.

• Strict-mode prompt-level: 85% → 90% (+5pp)
• Strict-mode instruction-level: 86% → 90% (+4pp)
• The improvement mainly came from fewer formatting errors

In session cost estimates, as the cache prefix and conversation history grow together, both cache read/write costs and rate-limit consumption speed were calculated to increase. Under an assumption of an 80-round-trip session, 4.6 was estimated at about $6.65, while 4.7 was estimated at about $7.86–$8.76, meaning that even with the same per-model pricing, actual operating costs rise further.

The key conclusion is that this is a model where, in exchange for improved accuracy and finer instruction-following, you must accept additional token costs, especially in English- and code-centric workloads.