🔢 Claude Opus 4.7 changed its tokenizer — and that has cost implications
Simon Willison upgraded his token counting tool to compare multiple Claude models simultaneously.
📊 The key data:
| Model | Tokens (text) | Ratio |
|---|---|---|
| Claude Opus 4.7 | 7,335 | 1.46x |
| Claude Opus 4.6 | 5,039 | 1.00x |
💰 Pricing hasn’t changed — $5/M input tokens, $25/M output tokens. But with ~46% more tokens for the same text, the effective cost increases proportionally.
🖼️ For high-resolution images, the impact is even larger:
- 3456×2234 px image: 3.01x more tokens in Opus 4.7 vs 4.6
- (The increase is because 4.7 can process higher-resolution images, up to 2,576px on the long edge)
- Small images: practically the same cost
🛠️ The tool is free, accepts text and images, and supports all four current main models.
💡 Explanation in a nutshell#
A “token” is the minimum unit LLMs process — roughly ¾ of an English word. Models charge per token, so if the same text generates more tokens, the cost goes up. Opus 4.7 uses a new tokenizer that generates ~46% more tokens for the same text, which is important when estimating costs in production.
More information at the link 👇
