Insane cache hit

Ran my coding agent (Reasonix/OMP) for a week against a few different models. Overall: - 2,046 requests - 271.7M tokens - $4.71 total spend Per model: > Kimi-K2.6 - 846 requests - 88.3M tokens - 98.2% prompt cache hit - $1.68 total > GLM-5.2 - 392 requests - 104.8M tokens - 99.3% prompt cache hit - $2.85 total > DeepSeek-V4-Flash - 808 requests - 78.6M tokens - 86.1% prompt cache hit - $0.18 total The biggest takeaway is how much prompt caching matters. My agent keeps a massive repository in context, so almost every request reuses cached prompt tokens. That makes long-running coding sessions dramatically cheaper than the raw token counts would suggest. Curious to hear what cache hit rates other people are seeing with coding agents.