Token use in Claude can quickly add up
Every message you send costs more than the last. Claude re-reads your entire chat from the top, to answer every message.
100K
tokens are used in a 20-message Claude chat
2.2×
more tokens are used going from 20 to 30 messages in a Claude chat
98.5%
of tokens in long chats go to re-reading chat history, not output
1,500–3k
tokens used for a PDF page, vs ~200 for the same text as markdown
~1,300
tokens used for a full screenshot – whereas a tight crop can be under 100
5×
the token use if you upload the same file to 5 separate chats
The longer a chat runs, the more tokens each new message burns. Most of that is just re-processing what’s been said.
Habits that cut token waste
Shorter chats, smarter inputs, and a few simple habits make a big difference.