A preprint examines prompt detail, thinking effort and low-cost probes as predictors of agentic coding costs for unseen tasks in a benchmark.