“A 5-line config change cut our own classifier bill by 82%. Then we watched 6 FDE deployments do the same.”
Claude's prompt caching went GA in February 2025. It is one of the highest-leverage configuration changes available to anyone running Claude in production — and most teams still don't use it.
Quick mechanics: when you mark a prefix of your messages array as cacheable, Anthropic stores the KV cache for that prefix on their side for 5 minutes. The next call that re-uses the same prefix only pays for the new tokens. Cache writes cost 1.25× the normal rate; cache reads cost 0.1×.
The break-even point is two requests within five minutes. If your system prompt + tool definitions + few-shot examples are stable across requests (and they almost always are), you save money the moment any request hits a warm cache.
The pattern.
For the FDEOS implementer classifier, our system prompt is 4,200 tokens (defines the role taxonomy, the canonical capability slugs, the scoring rubric). Without caching, every classified job pays for 4,200 input tokens at Sonnet rates.
With cacheControl: { type: 'ephemeral' } on the system prompt, the first job in any 5-minute window pays 1.25× — every subsequent job pays 0.1×. At 50 jobs/min during ingest, that's a 91% cost reduction on the system prompt tokens.
Total bill impact for us: $40K/mo → $7K/mo for the same workload. We have not changed anything else.
Three patterns to cache in implementer work.
-
Stable system prompts. Role rubrics, taxonomy, evaluation criteria. If it doesn't change per request, cache it.
-
Tool definitions. Every tool call you give Claude is part of the input. If you're using the same 6 tools across 1,000 requests, cache them.
-
Long few-shot examples. If you're showing Claude 8 worked examples to ground its style, those are cacheable.
What not to cache.
The user-specific part of the request — the actual job description, the actual implementer profile, the actual customer query. Those change per request and are wasted cache writes.
The full configuration code for the AI SDK + Anthropic SDK + the LangSmith trace patterns + the eval harness we use to verify cache hits — below.
Get the next one in your inbox.
We don't publish often — when we do, it's because we found something worth sitting with.
- Market intelThe implementer stack, May 2026
Claude leads at 64% mention. Cursor is up. n8n is the dark horse. What the postings actually demand.
- Tool deep-diveVoice agents quietly took over enterprise intake
Three under-the-radar deployments in healthcare, plus the four tools every implementer now needs to know.