Every model reply reports how many tokens of each type it used. Tokenmeter multiplies each type by that model's list price and adds them up:
cost = cached reads × read price + uncached input × input price + cache writes × write price + output × output price, all per million tokens
The types are priced very differently, so the token total alone does not predict cost. On Claude a cached read is about 2% of a fresh token, and output is 2 to 5 times a fresh token. Claude Code writes its cache with the 1-hour rate. OpenAI models do not charge for cache writes. The rates in use right now:
These are API list prices from pricing.json. If you are on a subscription you pay the plan price instead; the Spend tile shows that first and this API-equivalent figure below it.
Cache reads are cheap, cache writes and output are not. Thinking tokens are included in output.