Client-side token counter vs API-reported counts: which should you trust?

How this is calculated

For OpenAI models, the tiktoken library or gpt-tokenizer npm package gives exact pre-request counts that match the API response. For Llama models, llama3-tokenizer-js gives exact counts. For Claude and Gemini, client-side tokenizers don't have access to the real tokenizer, so any pre-request count is an estimate. The API response always includes the exact count. For cost estimation, use client-side tools for budgeting before you send. For billing and monitoring, use the API response values. If you see a mismatch between your client-side count and the API response for OpenAI models, check that you're using the correct tokenizer for the model (o200k_base for GPT-5 and GPT-4o, cl100k_base for older models).

Verdict

The two values answer different questions. tiktoken and gpt-tokenizer (o200k_base for GPT-5, cl100k_base for older models) and llama3-tokenizer-js all match the API response exactly because they ship the same tokenizer. Claude and Gemini keep their tokenizers private, so any pre-request estimate lands within about 10%. For budgeting before you send, run the client-side counter. For billing and monitoring, the API response usage field is ground truth.