An agent's average latency and cost rise after an update although answer quality is unchanged. Which evidence should be inspected?

1. Per-request token use, tool-call counts, retries, and model selection
2. Only the application display name
3. Only the homepage's visitor count
4. Only the final answer's sentiment