Claude Opus 4.7 costs 20–30% more per session

Anthropic's latest flagship model, Claude Opus 4.7, has quietly introduced a significant change: a new tokenizer that uses substantially more tokens to represent the same content. According to detailed testing, the model uses 1.35x more tokens on average compared to version 4.6, with technical documentation seeing the biggest jump at 1.47x. Code-heavy content is particularly affected, with TypeScript using 1.36x more tokens and Python 1.29x more. This means users' context windows fill faster, cached prefixes become more expensive, and rate limits are hit sooner—all while paying the same per-token price.

Anthropic's migration guide suggests this change enables "more literal instruction following" and better tool-call precision, particularly for character-level tasks. The theory is that smaller, more granular tokens force the model's attention mechanism to focus on individual words rather than larger patterns. Early reports from partners like Notion, Warp, and Factory indicate fewer tool errors in long-running sessions. However, the 20-30% effective price increase raises questions about whether the improved performance justifies the additional cost, especially for developers working with code-heavy prompts where the token inflation is most pronounced.