Opus 4.7's New Tokenizer Inflates Text Costs ~40% Despite Unchanged Pricing

Simon Willison extended his Claude Token Counter to compare token usage across models, surfacing a practical consequence of Opus 4.7’s updated tokenizer. Anthropic’s own announcement flagged a 1.0-1.35x input inflation, but Willison’s test on the Opus 4.7 system prompt came in at 1.46x against Opus 4.6. With pricing held flat at $5/$25 per million input/output tokens, that translates to roughly 40% higher real-world spend for equivalent text workloads.

Images and documents tell a more nuanced story. A high-resolution PNG produced a 3.01x token count on 4.7, but that gap is driven by Opus 4.7’s expanded vision support for images up to ~3.75 megapixels - at comparable resolutions, token counts are effectively identical. A 30-page text-heavy PDF landed at just 1.08x, suggesting the tokenizer penalty is concentrated in raw prose rather than mixed-media inputs.

The takeaway for teams budgeting against Claude: tokenizer changes are a hidden price lever. Nominal per-token pricing stayed flat across the 4.6-to-4.7 jump, but actual cost-per-task shifts materially depending on input type, and anyone forecasting spend should re-benchmark against representative payloads rather than trusting headline rates.