There is no retrieval method for assessing tokens used. Costs are even obfuscated in the usage page of your account.
Assistants are loaded with tokens you didn’t write, loaded to the max with conversation and retrieval, and can make iterative calls to internal functions.
All language received and emitted by the AI model behind assistants framework uses billable tokens. Instructions are part of what is placed into the AI context.
A “run” of a thread can have a new instruction attached that replaces the assistant’s instruction. “Threads” does not allow any messages to be added to the conversation other than user role messages.
I hope that lets you determine if “assistants” is still an acceptable way for you to interact with OpenAI models.