Developed prompt caching for Anthropic models in the zed-industries/zed repository, focusing on efficient API integration and backend development using Rust. The implementation introduced explicit per-block cache breakpoints and session ID sticky routing, ensuring accurate token usage accounting and compatibility across Anthropic, Bedrock, and Vertex AI providers. By avoiding reliance on OpenRouter’s top-level cache control, the solution maintained consistent routing and cache behavior. Comprehensive tests were added to validate caching logic across multiple model types, improving observability and predictability of costs. This work reduced latency for cached prompts and enhanced the reliability of prompt engineering workflows involving large language models.