Myth: if the price per million tokens stays the same, your bill stays the same. Fact: the number of tokens your text produces depends on the model's tokenizer, so a new model can turn the same prompts into more tokens. Recount your own prompts on the target model before you estimate cost.
What the documentation says
The Claude token counting page states that models from a certain generation onward use a newer tokenizer, and the same input text produces approximately 30 percent more tokens than on earlier models. It adds that the exact increase depends on the content and workload shape, and tells readers to recount prompts against the model they plan to use rather than reuse earlier counts.
A later section says usage and billing on those models reflect the new tokenizer's counts, and that you should not reuse counts from an older model to estimate cost or context window fit.
That figure is the vendor's statement for one comparison. Your number could be lower or higher.
Three more places the myth breaks
Output behavior. Models differ in how verbose they are and how much they reason. Same price, different output tokens.
Defaults. Effort defaults differ by model. The effort docs say two newer models default to medium while others default to high, and one note says a request that omits the setting runs one level lower than it did on the previous model. Omitting a parameter can change spend.
Caching rules. Minimum cacheable lengths and read multipliers differ by model in the caching tables. A prompt that cached on one model may fall under the minimum on another, or the reverse.
Test it in an afternoon
- Collect a few hundred real requests, including system prompts and tools.
- Run the token counting endpoint for each against the current and candidate models. It is free, with a requests per minute limit.
- Compare the totals by feature.
- Run your evaluation set on both models and record real output and reasoning tokens from the usage fields.
- Compute cost per completed task for each, with caching assumptions stated.
What to do with the result
If the candidate model costs more per task but passes more tasks, the extra cost may be worth it. If it costs more and quality is equal, adjust effort, prompts, or stay put. Either way you decide with data.
Budget in dollars as well as tokens
If you set token limits in config, such as a maximum prompt size, they were calibrated to one tokenizer. Revisit them at every model change, along with your context window headroom.
Quick check
Is the 30 percent figure a guarantee?
No. The docs say the exact increase depends on content and workload, and describe it as approximate. Measure your own prompts.
Do I need to rewrite prompts for a new tokenizer?
Not necessarily. You need to recount them and re-run your evaluation, because both cost and behavior can shift.
Sources
Last reviewed: 19 October 2026.
