Prompt caching, explained: why your second agent request costs less

What prompt caching is

Every time an agent takes a step, it sends the model the whole conversation so far: your rules, the files it has read, the edits it has made and the results of any commands it ran. Most of that is identical to what it sent a few seconds earlier.

Prompt caching lets the model provider recognise that repeated beginning of the request and skip reprocessing it. Because it's cheaper for them, providers charge less for it.

How cached tokens are priced

Each provider sets its own prices, but the pattern is similar:

• Cached input is billed at a fraction of the normal input price, often around a tenth.
• Writing to the cache is free with some providers and costs a small premium over normal input with others.
• Output tokens are never cached and are always billed at the normal output price.

Flintr bills all three at the provider's list price, so any saving from caching goes straight to you.

How it shows up on your bill

Open the cost breakdown on any agent step and you'll see cached tokens listed separately from fresh input:

Step 6 · Claude Sonnet
input     2.1k   fresh
cached   16.3k   read from cache
output    0.9k
cost     $0.03

In this example, the step read almost 90% of its input from the cache, which costs a fraction of sending it fresh.

Habits that get more cache hits

• Keep conversations focused. Start a new chat for a new task instead of piling unrelated work onto a long thread.
• Keep rules stable. Rules sit at the very start of every request. Editing them mid-task means the cache has to start over.
• Avoid editing early messages. Changing something near the top of a conversation invalidates everything cached after it.
• Stay on one model during a task. Caches belong to a single model, so switching models resets them.