QUICK ANSWER
GPT Token Counter
A GPT token counter must use the encoding associated with the exact GPT model. Paste text to measure its encoded tokens before a request, but use the input and output usage returned by the OpenAI API for the final count of a completed call.
GPT models can use different encodings
OpenAI assigns encodings to model families. The same text can split differently under different encodings, so selecting a generic GPT label without identifying the target model may not produce the count used by that model.
A pasted-text count has limits
A local counter can measure the text it receives. An API request may also include system instructions, conversation history, message formatting, tool definitions, images, audio, or other content that is not present in the pasted text.
Separate counting from pricing
Tokenization determines the amount of usage. Pricing then applies the selected model’s current input, cached-input, and output rates where those categories are published. Changing the model can change both the token count and the price.
Use the matching page
Use the OpenAI Token Counter to count visible text with a supported OpenAI encoding. Use the Token Cost Kit homepage to compare token-based model costs. Read Counter Token when you need to distinguish AI tokens from other uses of the term.