Claude Opus 4.7 Costs 40% More Than 4.6 — Same Price, Different Tokenizer
Simon Willison's upgraded token counter tool reveals that Opus 4.7's new tokenizer inflates counts by 1.46x over Opus 4.6 on the same input — far above Anthropic's stated 1.0–1.35x range. At identical list pricing of $5 per million input tokens, that hidden inflation makes 4.7 effectively ~40% more expensive to run in practice.
Original sourceSimon Willison published an upgraded token counting tool today that compares tokenization across four current Claude models: Opus 4.7, Opus 4.6, Sonnet 4.6, and Haiku 4.5. The headline finding is uncomfortable for anyone who upgraded to Opus 4.7 expecting equivalent costs.
**The tokenizer gap is bigger than Anthropic advertised.** When Willison ran Opus 4.7's own system prompt through the comparison tool, 4.7 consumed 1.46x more tokens than 4.6 on identical input. Anthropic's stated range for tokenizer differences was "1.0–1.35× depending on content type" — the real-world number Willison measured sits well above the ceiling of that range.
**What this means at the pricing level.** Opus 4.7 and 4.6 share the same list price: $5 per million input tokens. But if you're sending 100M input tokens a month and those tokens inflate by 1.46x on 4.7, you're actually paying for 146M tokens — an effective cost increase of about 40% with zero pricing change disclosed.
**It's not uniform across content types.** High-resolution images showed the most dramatic differences (3.01x more tokens on 4.7), but Willison clarified this is entirely attributable to 4.7's superior resolution support — the same image processed at equivalent resolution shows near-identical counts. PDFs showed only 1.08x inflation. The worst-case inflation appears to be raw text.
**The tool is public and functional.** Developers running cost-sensitive Opus deployments can now run their own inputs through the comparison tool to get a concrete multiplier before committing to a model upgrade. The tool supports all four current Claude models and outputs side-by-side token counts in real time.
This is the second time in two months that a community-built tool has surfaced a meaningful Claude cost discrepancy that Anthropic's official documentation understated. The pattern suggests that API pricing transparency for tokenization changes deserves more scrutiny across the industry.
Panel Takes
The Builder
Developer Perspective
“Every team running Opus at scale needs to run Willison's tool on their actual production inputs before upgrading to 4.7. A 40% effective cost increase with no pricing change on paper is a budget planning problem. Anthropic should publish per-model tokenizer coefficients in the docs, not bury them in blog post ranges.”
The Skeptic
Reality Check
“The 1.46x figure is real but the framing is selective — if 4.7 is genuinely better at 4.7's tasks, the cost-per-correct-output might actually be lower even with more tokens. The image inflation turns out to be a feature (higher resolution processing). The honest question is whether you need 4.7's capabilities or are paying a premium for marginal gains.”
The Futurist
Big Picture
“This is what model versioning opacity looks like at scale. As AI becomes infrastructure, hidden tokenizer changes that shift effective costs 30-40% are governance failures, not just docs issues. The community building comparison tools to surface what providers should disclose proactively is a sign of market maturity — and of how much trust still needs to be built.”