Anthropic's Claude Opus 4.7 Tokenizer Change Quietly Raises Real-World Costs Up to 27% Despite Unchanged Pricing
Summary
Anthropic's Claude Opus 4.7 quietly raises real-world costs up to 27% through a new tokenizer that generates 32–45% more tokens than its predecessor, despite keeping prices unchanged at $5/M input and $25/M output — though short prompts under 2K tokens actually see a slight cost decrease.
Key Points
- Anthropic's Claude Opus 4.7 introduces a new tokenizer that generates 32–45% more native tokens than Opus 4.6 for equivalent text, meaning costs rise despite the model price remaining unchanged at $5/M input and $25/M output.
- Real-world usage data from over one million OpenRouter requests reveals that actual costs increased 12–27% for prompts above 2K tokens, with prompt caching absorbing a significant portion of the inflation — up to 93% for the longest prompts exceeding 128K tokens.
- Short prompts under 2K tokens are bucking the trend, seeing a 1.6% cost decrease as Opus 4.7 produces 62% fewer completion tokens for simple queries, fully offsetting the tokenizer overhead.