A token is a model unit
A tokenizer splits text into units that can be words, word fragments, punctuation or spaces depending on the model. “Unbelievable” may become multiple tokens while a common short word may be one. Code, numbers and languages other than English often produce a different ratio from English prose.
Why estimates differ
A words-to-tokens estimate is useful before a request, but it is never a substitute for measured usage. Providers use their own tokenizers and may account for message structure, system instructions, images or tool definitions differently. A one-thousand-word English draft is often several hundred more tokens than its word count, but the exact number depends on content and model.
Input and output have separate roles
Input tokens include what you send, including repeated history. Output tokens cover what the model generates; reasoning models may count additional internal work as output. The API usage object is the best source for these counts. Convert your observed token counts to dollars with the model's rates.
A simple measurement workflow
Save a representative set of requests. Record model ID, input tokens, output tokens and any cached usage reported. Calculate cost per request and compare it with your provider dashboard. When counts disagree, check whether the response included tool use, images, reasoning or a service tier you left out.