OpenRouter adds response caching for identical requests

Hot take: this is a practical infra upgrade that matters more than it sounds, because repeated LLM calls are one of the easiest ways to burn tokens and time.

• –It directly cuts cost on deterministic or repeat-heavy workloads.
• –The feature is developer-friendly: a header opt-in and explicit cache status headers make behavior observable.
• –This is especially useful for eval loops, regression tests, and prompt experiments.
• –The main limitation is that caching only helps when requests are truly identical, so it won’t change fundamentally dynamic chat workloads.