The cost nobody counts

Flip each read-only item from Tool to Resource and watch both meters. One of them you can see in a token counter. The other you can't, until something breaks.

Context cost / turn
960
tokens, illustrative · ~120/tool schema
Decisions handed to the model
4
read-only items still gambling
Tool: schema loaded every turn Resource: fetched only when needed
read_page TOOL
list_pages TOOL
get_customer TOOL
search_docs TOOL
write_page TOOL a write is a real decision; it stays a Tool
delete_page TOOL a write is a real decision; it stays a Tool
What each meter shows. The token count is the cost you see: it's right there in your usage dashboard. The decision count is the cost you don't: it's the number of times, per turn, the model has to guess whether to fetch data it obviously needs. A token counter would never show you the run where it just answered, confidently, about data it never fetched.
Same behavior. Fewer tokens. Zero coin-flips. The token bill was the cheap part.