Comment by lynox

hey danilo, i just recently fixed a bug which broke caching in my agent loop. the cost explosion was immense: on my canary instance i paid like $17/day instead of about $4. i just recently measured anthropic and mistral model cost and found that without caching the cost is about 2-4x higher in general. in my case the fix was software side so i cant help you with the google support. i noticed cache_read-tokens per turn were around 0 cause the prefix invalidated every turns cache.