A team uses a model priced separately by input and output tokens. It repeatedly sends a very long prompt and requests lengthy answers. Without changing the number of requests, which change can directly reduce token-based inference usage while preserving required content?

1. Keep sending the same oversized prompts and expect a lower temperature to eliminate input-token charges.
2. Count only the number of characters and assume every model uses one character per token.
3. Increase the requested answer length because output tokens are always free.
4. Remove unnecessary prompt material and set an appropriate response-length limit, then evaluate answer quality.