Simon Willison · 2026-04-20 · notable
Simon Willison: Claude Token Counter with Model Comparisons — Opus 4.7 Tokenizer Costs ~40% More
Simon Willison upgrades his Claude token counter to compare models side-by-side and finds Opus 4.7's new tokenizer uses 1.46x more text tokens than Opus 4.6 — identical per-token pricing, so ~40% more spend per typical prompt.
Opus 4.7's new tokenizer produces 1.46x more text tokens than 4.6 — same price per token, but effectively ~40% more spend per prompt.
What is it?
A blog post and tool update from Simon Willison, published April 20, 2026. Willison extended his Claude Token Counter web app to show token counts for multiple Claude models side-by-side, then ran real prompts through it and surfaced a practical pricing implication: Opus 4.7's new tokenizer produces significantly more tokens from the same text than Opus 4.6.
How does it work?
The upgraded tool calls Anthropic's token counting API against each model in parallel and shows the results side-by-side. The key measurement: the same system prompt consumed 1.46x more tokens on Opus 4.7 than 4.6 — above Anthropic's own stated 1.0–1.35x range for the tokenizer change. A high-resolution PNG showed a 3.01x ratio, but this was entirely due to Opus 4.7's new ability to process images at up to 2,576 pixels (not pure tokenization overhead). PDFs showed a moderate 1.08x delta.
Why does it matter?
Anthropic held Opus 4.7 pricing flat at $5/M input tokens, so the new tokenizer is an effective hidden cost increase for text workloads: expect roughly 40% more input spend when migrating from Opus 4.6. The tool is now a practical pre-migration audit — paste your actual system prompt and see the token delta before committing.
Who is it for?
Teams budgeting API costs for a migration from Claude Opus 4.6 to Opus 4.7.
Try it
https://tools.simonwillison.net/claude-token-counter