A model produces correct short answers but is prompted to explain every simple extraction in several paragraphs. Cost and completion latency are excessive. Which change should be tested first?

1. Use a concise task-specific response contract and appropriate output token limits.
2. Increase the maximum output tokens for every request.
3. Remove the required extraction fields from the answer.
4. Use only a longer client timeout.