Claude's Newer Models Tokenize ~30% Denser — and Most Cost Calculators Haven't Noticed

Buried in Anthropic's model documentation is a line with real budget consequences: models from Claude Opus 4.7 onward use a new tokenizer that produces roughly 30% more tokens for the same text. Their own context-window figures say it quantitatively — a 1M-token window is documented as ~2.5M characters for the new tokenizer against ~3.4M for the previous one.

The same posted price per million tokens therefore buys meaningfully less text on the newer models. At $5/M input, 40,000 characters of prose is ~11,800 tokens on Claude Sonnet 4.5 and ~16,000 tokens on Opus 4.7 — a ~36% difference in what you are billed, for identical input, at an identical sticker price.

The trap inside the trap

The obvious fix — "multiply Claude 4.7+ by 1.3" — has two failure modes we hit while shipping it:

Membership is not a version comparison. Claude Sonnet 5 uses the new tokenizer. Claude Sonnet 4.6, Opus 4.5, and Haiku 4.5 use the previous one. No >= 4.7 check gets that right; the mapping has to come from the per-model figures Anthropic actually publishes.

A flat base ratio was also wrong. Our first fix applied 1.3× over the generic 4-chars/token baseline. But Anthropic's previous tokenizer runs ~3.4 chars/token, not 4 — so Sonnet 4.5 had been silently under-estimated by ~18% all along, hidden by the fact that the generic baseline happened to be the industry default rather than the vendor's number.

Where our numbers come from now

Every characters-per-token ratio in the calculator carries an evidence label: primary when the vendor states it (OpenAI and Gemini both publish ~4; DeepSeek publishes its own figures), derived when computed from vendor-published figures with the arithmetic shown (Anthropic's 2.5 and 3.4), measured when we ran our public corpus through the vendor's own tokenizer, and assumed when nobody publishes anything — labelled as such on the page rather than dressed up as fact.

If a cost tool shows you one ratio for every model, it is wrong by a third on some of the most heavily used models in the industry. Ask it where its number comes from.