Which start of the prompt is reused?

Earlier request

The words [document] stand for a long pasted text, counted as stand-in tokens.

Mark: after [document], else before the last line.

New request, split into tokens

A real model splits text differently. Prices are examples. Only input tokens are shown; output is billed as usual.

More settings

Rules A to D are the four rules listed, with their sources, in the lesson. A rule can change at any time.

What this toy leaves out

A server may drop a stored copy earlier than its lifetime, and two requests may not reach the same copy. Anthropic also looks back up to 20 blocks before a mark and allows up to four marks. Images and tools follow their own rules. Real prices change.