RealFeedApp Kompressor reduced costs by 22.7%
In Claude Code, 92–95% of tokens come from a cache costing $5 for 5 minutes, while reading tokens from cache costs $0.20 per million. This creates a bottleneck for agents needing to wait for cache expiration before executing tasks.
- Field
- applying language models
- What they did
- The company Hystax discovered that accessing the cache in Claude Code costs 25 times more than reading from it, and that the cache expires after only five minutes.
- Why it matters
- This knowledge will help developers optimize costs when using Claude Code, understanding that caching is inefficient during long agent cycles.
#claude code#api billing#cache#llm#cost optimization
