Cache tokens

Cache tokens are the portion of a request's input tokens that the model provider recognised from its prefix cache and didn't have to process again. They still count toward the context window (the model sees them) but are billed at a fraction of the normal input rate and cost almost no latency. On a healthy coding session, the large majority of input on each request should be cache tokens.

They're the number to watch when a session feels expensive. If a usage readout shows a big context but few cache tokens, something is breaking the cache on every request: an instruction file that's regenerated each turn, a tool whose definition changes, or gaps between messages long enough for the cache to expire. If cache tokens are high and the bill is still climbing, the problem is simply that the window is too big, and compaction is the fix.

Providers report them separately, and harnesses surface them. Learning to read that line ('cached' vs 'fresh' input) turns cost from a mystery into a diagnostic. Cache-write tokens are the other side: the first time a prefix is stored, some providers charge a small premium. That's fine as long as it's followed by many reads.