Crowkis vs Anthropic prompt caching: cache writes that bill you are telling you something
Anthropic's prompt caching is excellent at its actual job, cheap long contexts. It was never designed to be your response cache, and the pricing says so.
Anthropic's prompt caching shines for what it was built for: agents and chat apps re-sending a large stable context get the repeated prefix at a steep discount, with cache writes billed at a premium and entries expiring in minutes. Take that deal for your long system prompts, we do. But read the design closely and it tells you its scope: short-lived, prefix-exact, input-side, single-provider.
Five-minute TTLs mean your cache evaporates between user sessions. Prefix-exactness means paraphrased questions never qualify. Input-side means the answer, the expensive part, is regenerated every time. And provider-side means your accumulated cache value is a feature of someone else's pricing page, revocable and non-portable.
- 1your app · redis-py · ioredis · Lettuce
- 2crowkis
- 3claude code · agents
- 4services
- 5your LLM provider
Four doors in, one cache, and the model only sees genuinely new questions.
Crowkis is the durable, output-side complement: answers persist on your disk for as long as your TTL policy says, match by meaning rather than prefix, survive restarts and model swaps, and serve in under a millisecond without touching the API at all. Your cache becomes an asset you own rather than a discount you receive.
The bottom line
The composition is clean: Anthropic's caching cuts the cost of calls Crowkis decides must happen; Crowkis eliminates the rest. One optimizes the pipe, the other installs a reservoir.