Claude Opus 4.7 and the Cost of Token Inflation

Simon Willison's updated token counter tool reveals that Claude Opus 4.7's new tokenizer significantly increases token counts for the same input compared to previous models. This 'token inflation' means that while per-token pricing is unchanged, the actual cost of processing text and high-resolution images has risen by 40% to 300%. The impact varies by content type, with PDFs seeing a much smaller increase than raw text or large images.

Key Points

• Simon Willison updated his token counting tool to compare Claude models, revealing significant tokenization changes in Opus 4.7.
• Claude Opus 4.7 uses a new tokenizer that increases token counts for the same text by approximately 46%, leading to higher effective costs.
• The model's improved vision support allows for higher-resolution images, which can result in a 3x increase in token usage for large image files.
• Token inflation is inconsistent across formats, with text-heavy PDFs showing only a 1.08x increase compared to the much higher multipliers for raw text and high-res images.

Sentiment

The community leans toward agreeing that token inflation is a real and concerning issue, but is meaningfully divided on intent. The majority view treats the measurable cost increase as problematic and criticizes Anthropic's silence about it. A significant technically-informed minority provides plausible defenses and argues the conspiracy framing doesn't hold up economically. A pragmatic middle ground wants to see cost-per-task data before rendering judgment. The overall tone is critical but not hostile — more demanding accountability than calling for boycotts.

In Agreement

• The 1.46x token increase is confirmed independently by multiple users and exceeds Anthropic's stated range, making the model effectively 40% more expensive
• Subscription tier limits burn through faster, forcing users toward more expensive plans — a pattern also seen at other AI providers
• Anthropic's lack of transparency about why the tokenizer changed fuels distrust, with the secret tokenizer and hidden thinking tokens contributing to a perception of an increasingly closed ecosystem
• Token-based pricing is inherently unfair across languages, with English speakers paying less than Spanish speakers for equivalent content
• Agent workflows compound the cost increase due to retry loops that re-send full context, potentially tripling token costs on failures
• Several users are actively switching back to Opus 4.6, trying local alternatives like Qwen 3.6, or exploring competitors like Codex

Opposed

• More tokens cost Anthropic more compute to process, so a tokenizer-based money grab makes no economic sense — they could simply raise per-token prices if they wanted more revenue
• The tokenizer change likely has legitimate technical justifications, possibly a more semantically aware tokenizer or the learned Mythos tokenizer that improves model quality
• Early benchmarks suggest Opus 4.7 may arrive at answers with fewer total tokens despite the per-token inflation, making cost-per-task potentially equivalent or better
• Anthropic can't keep up with demand already — companies in that situation just raise prices directly rather than engineering sneaky tokenizer schemes
• The real metric should be cost per successful task completion, not cost per token — tokenizer data alone is incomplete for evaluating value
• Speculative decoding could make the additional whitespace tokens trivially cheap to generate, meaning Anthropic's actual compute costs may not increase proportionally