Prompt Caching

Prompt caching stores the stable part of what you send to the model, typically your instructions, tool definitions and reference documents, so repeat calls read it at around a tenth of the normal input price. It is plumbing rather than a feature, and it routinely takes more than half off a production bill.

Cached reads cost about ten per cent of the normal input price, and real workloads typically see their input bill drop by half or more. The exact number depends on how much of your prompt repeats, which is what we measure first.

No. The model sees exactly the same context either way. Caching changes what you pay to send it, not what Claude does with it.

Usually a cache breaker: changing content placed before stable content, a timestamp in the prompt, or calls spaced past the cache lifetime. These are design faults, and they are fixable.

If a workload repeats context, usually yes, because the write premium pays back on the first reuse. For low, irregular traffic it may not matter, and we will tell you that plainly.

The parts that do not change between calls. That is usually your system instructions, the tool definitions and any reference material the model needs every time. The user’s actual question is the part that varies and is not cached.

Not long, which is why it suits workloads with steady traffic rather than occasional use. A busy production agent keeps the cache warm continuously. Something called twice a day will mostly miss it.

More than anywhere else. An agent makes many calls carrying the same instructions and tools each time, so it is the workload where caching makes the largest difference. Building an agent without it is the most common avoidable cost we see.