A customer asks why their bill is 30% higher after moving to a newer model with the same prompts. What is going on, and how do you reason about tokens?
Answer it out loud first, then compare. That's what makes it stick.
Models read and bill in tokens, not words, and each model family has its own tokenizer. Anthropic documents that models from Opus 4.7 onward use a newer tokenizer that produces roughly 30 percent more tokens for the same text. So identical prompts cost more and fill the context window faster. I would recount the real prompts with the token counting endpoint on the target model, then rebudget cost and context from those numbers, not from old estimates.
- Tokens, not words, are the unit of cost, limits and latency
- Tokenizer differs per model generation
- Recount with the count-tokens endpoint on the target model
- Separate input, output and thinking tokens in the bill
- Rebudget cost and context headroom, then cut waste
A token is a chunk of text (a word piece, punctuation, whitespace) that the model's tokenizer produces. Everything that matters operationally is counted in tokens: price (per million tokens), context limit, rate limits, and a large share of latency (output tokens are generated one at a time).
What changed here. Anthropic's token counting docs state that Claude 4.7 and later models use a newer tokenizer, and the same input text yields about 30 percent more tokens than earlier models, with the exact increase depending on content. The models overview adds that 1M tokens is roughly 555k words on the current tokenizer versus about 750k words before. So a customer who copies old token estimates will under-budget both money and context.
How I would handle it with the customer:
- Take 50-100 representative production requests and call the count-tokens endpoint (free, rate-limited per minute, an estimate that can differ slightly from billed usage) once on the old model and once on the new one. Compare the two numbers.
- Read the real
usageobject on live traffic. Break the bill into input, output, cache write and cache read tokens. Thinking tokens bill as output. - Find the fat: long system prompts, verbose tool definitions, full documents pasted when a section would do, chat history that never gets trimmed.
Levers, in order of effort:
- Prompt caching for the stable prefix (reads cost a fraction of base input).
- Trim tool definitions and boilerplate.
- Cap
max_tokensand ask for concise output, since output tokens cost several times more than input. - Route easy traffic to a smaller model.
- Batch API for non-urgent work (50 percent off).
Numbers to quote. On Sonnet 5.5 list prices of $2 input and $10 output per million tokens, a 3,000-token input with a 300-token output costs about $0.009 per call. At 100k calls a day that is roughly $900, so a 30 percent token increase is a visible five-figure monthly line.
Close by saying you would put token counts per request type on the customer's dashboard so the next model migration surprises nobody.
- How would you estimate cost for a new customer before you have any traffic?
- Why is output more expensive than input, and what does that mean for prompt design?
- How do images and PDFs count against the budget?
- Token counting - Claude docs (docs, 2026-11) Count-tokens endpoint is free, an estimate, and Claude 4.7+ models produce about 30 percent more tokens for the same text.
- Models overview - Claude docs (docs, 2026-11) Current per-million-token prices, and 1M tokens is about 555k words on the current tokenizer.