The reality is that it’s not easy to estimate your cotst on Assistants, not even by tracking your messages.
The only way is by controlling all run steps and counting the tokens from each and every event and message.
If you need predictability, you need a different stack. The assistant API is in beta and should not be used for production applications.
Other users have the same issue. We’d love to get that predictability, but no word from OpenAI yet.