Capture the usage field (input_tokens / output_tokens) and model id on each answer, and total them in the report as a TOKEN COST line with a per-account average. Makes the run's actual cost visible alongside accuracy, so the option set and sample size can be traded against spend from measured numbers rather than a guess. First real call confirmed the response shape (answers.country.{choice, probabilities,confidence}) needs no parser change; model jev-1.13.0 returned ~1736 input / 339 output tokens per account over 45 country options.