Simon Willison's token counter shows Opus 4.7 is 40% more expensive
Simon Willison shipped a model-comparison update to his Claude Token Counter today and it buries one of the quieter Opus 4.7 stories. The new Anthropic tokenizer charges 1.46× the tokens of 4.6 for the same system prompt. Anthropic told everyone 1.0 to 1.35×, depending on content type. Actual number is north of that.
Per-token pricing didn't change. $5 per million input tokens, same as 4.6. But your bill goes up 40% on typical workloads just because the tokenizer got more verbose. Code-heavy prompts take the worst hit. PDFs are only 1.08×. Images look 3× worse but Willison argues that's actually higher resolution at the same effective cost, not a regression.
The tool lets you paste text and see the split across every Anthropic model side by side. If you're running Claude at any real volume this is the kind of thing your finance team should look at before you auto-upgrade. A lot of teams moved to 4.7 for the capability jump and didn't price in the tokenization overhead.
Check it out at https://tools.simonwillison.net/claude-token-counter
← Back to all articles
Per-token pricing didn't change. $5 per million input tokens, same as 4.6. But your bill goes up 40% on typical workloads just because the tokenizer got more verbose. Code-heavy prompts take the worst hit. PDFs are only 1.08×. Images look 3× worse but Willison argues that's actually higher resolution at the same effective cost, not a regression.
The tool lets you paste text and see the split across every Anthropic model side by side. If you're running Claude at any real volume this is the kind of thing your finance team should look at before you auto-upgrade. A lot of teams moved to 4.7 for the capability jump and didn't price in the tokenization overhead.
Check it out at https://tools.simonwillison.net/claude-token-counter
Comments