Claude Token Counter gains model comparison for Opus 4.7 vs 4.6
Simon Willison upgraded his Claude Token Counter tool to compare token counts across models, revealing that Opus 4.7's new tokenizer can cost roughly 40% more than Opus 4.6 for text inputs despite identical pricing.
Score breakdown
Developers budgeting for Claude Opus 4.7 should account for up to ~40% higher costs on text workloads due to tokenizer inflation, and should test their specific content types — PDFs, images, and raw text behave very differently — using the updated token counter tool before migrating from Opus 4.6.
- 01The Claude Token Counter tool was updated to compare token counts across Opus 4.7, Opus 4.6, Sonnet 4.6, and Haiku 4.5.
- 02Opus 4.7 is the first Claude model to use an updated tokenizer, per the post.
- 03Anthropic states Opus 4.7's tokenizer maps the same input to roughly 1.0–1.35× more tokens depending on content type.
The post describes an upgrade to the Claude Token Counter tool that adds the ability to compare token counts across different Claude models side by side. The motivation is Opus 4.7, which the post identifies as the first Claude model to introduce a changed tokenizer. The tool now supports four models: Opus 4.7, Opus 4.6, Sonnet 4.6, and Haiku 4.5. Anthropic's own announcement for Opus 4.7 stated that its updated tokenizer "improves how the model processes text" but that the same input can map to "roughly 1.0–1.35× depending on the content type."
For a large 3456×2234 pixel, 3.7MB PNG image, the token count was 3.01x higher on Opus 4.7, but a follow-up test with a small 682×318 pixel image showed nearly identical counts (314 vs.
Testing with the Opus 4.7 system prompt yielded a 1.46x token count increase over Opus 4.6 — above Anthropic's stated range. Since both Opus 4.7 and 4.6 are priced identically at $5 per million input tokens and $25 per million output tokens, this token inflation means Opus 4.7 can be expected to run approximately 40% more expensive in practice. For a large 3456×2234 pixel, 3.7MB PNG image, the token count was 3.01x higher on Opus 4.7, but a follow-up test with a small 682×318 pixel image showed nearly identical counts (314 vs. 310 tokens), confirming the large-image difference stems from Opus 4.7's expanded high-resolution image support (up to 2,576 pixels on the long edge, more than three times prior Claude models) rather than tokenizer overhead alone. A 15MB, 30-page text-heavy PDF showed a more modest 1.08x multiplier (60,934 tokens on 4.7 vs. 56,482 on 4.6), suggesting the tokenizer impact varies significantly by content type.
Key facts
- 01The Claude Token Counter tool was updated to compare token counts across Opus 4.7, Opus 4.6, Sonnet 4.6, and Haiku 4.5.
- 02Opus 4.7 is the first Claude model to use an updated tokenizer, per the post.
- 03Anthropic states Opus 4.7's tokenizer maps the same input to roughly 1.0–1.35× more tokens depending on content type.
- 04Testing the Opus 4.7 system prompt showed a 1.46x token increase over Opus 4.6.
- 05Opus 4.7 and Opus 4.6 share identical pricing: $5 per million input tokens and $25 per million output tokens.
- 06A 3456×2234 pixel image produced 3.01x more tokens on Opus 4.7, but a small 682×318 pixel image showed nearly identical counts (314 vs. 310), attributing the large-image difference to Opus 4.7's expanded high-resolution image support.
- 07