Anthropic's Claude Opus 4.7 quietly raised the real cost of using Claude Code by 20-30%, and they didn't change the sticker price to do it. The culprit is a new tokenizer that breaks English and code into smaller pieces. Abhishek Ray ran measurements using Anthropic's own token counting API and found that real-world content like CLAUDE.md files and technical documentation uses 1.3x to 1.47x more tokens than the same content did on Claude Opus 4.6. Anthropic's migration guide claimed a range of 1.0 to 1.35x. The actual numbers for English-heavy content sit at the top of that range or above it.
Chars-per-token for English prose dropped from 4.33 to 3.60. TypeScript went from 3.66 down to 2.69. Your context window fills faster. Your rate limit hits sooner. AI Flooded One Firm With 1 Million Lines of Unreviewed Code shows how quickly token usage and output volume can spike, straining even generous rate limits. Meanwhile, GPT-4's tokenizer sits around 4.0 chars-per-token for English, and Llama 3 hits 4.0 to 4.5. Claude 4.7 is now the least token-efficient model in its weight class.
Anthropic's argument is that smaller tokens force better attention to individual words, which improves instruction following. Ray tested this with IFEval, a benchmark with verifiable constraints. On a 20-prompt sample, 4.7 scored +5 percentage points better on strict instruction following over 4.6. That's a real improvement. It's also a small one, measured on a small sample, and it could come from weight changes or post-training rather than the tokenizer itself.
The migration guide's optimistic range isn't a rounding error. It's the difference between 'you might not notice' and 'your bill just jumped by a third.' Anthropic had the real numbers. They ran the tokenizer on their own benchmarks before shipping. Understating the impact means fewer users hesitate before upgrading. Opus 4.6 and Sonnet 4.6 are still available if you want to vote with your wallet.