Configuring Token Count Metrics for Online Evaluators in AX Using Response Data Instead of Trace Estimates

hey guys, having some issues when getting a sense of costs running online evaluators with AX. we have set up some evals to run on prod traces like "hallucination, user frustration" etc. We are saving the traces separately into a project, but all the eval runs are showing an estimated token count of 4 chars per token. when we checked the attributes of the LLM traces in the project we're saving the eval traces to, we see the token counts similar to the estimates. Is there a way we can configure it to use the token counts from the responses/completions/external APIs instead? thanks