Token Usage & Cost

You might say

My API bill blew up this month. What am I actually paying for?

The cumulative model tokens in a task and the price calculated from the model’s rate cardSystem rules, chat history, files, tool results, and model output all use tokens across a multi-turn task. Cost also depends on the model and separate input, output, and cached-token prices; external tools may have their own fees.

Know first

Token Context Window

Why three turns cost more than the last answer alone

These rounds and Token counts are illustrative.

Total through three rounds1,380 TokenInput 1,200 + Output 180

Every round brings along needed rules, history, and new results; cost is not just the final answer.

How multi-turn usage and cost add up

Every turn has input: Rules, chat history, files, and tool results may enter the model again each turn.

Every turn has output: Responses, next-step suggestions, and tool requests also produce output tokens.

Calculate each token category separately: Add input, output, cache, and model categories, then apply the matching rates.

Quick checkChoose the best answer

An Agent resends long system rules, full chat, and tool results every round, and the bill keeps rising. How should cost be diagnosed and reduced?

Hide old messages in the chat interface and assume request cost falls automaticallyTrack input, output, cache, and model price per round, then remove irrelevant contextCount only the user's new sentences and ignore system rules, history, and tool results

You can say this to an AI Agent

Report each Agent round's input, output, cached tokens, model, and estimated cost, marking repeated history and large tool results. Remove irrelevant context before considering a cheaper model, and never count only new user text.

Learn next

1Context Window→2Rate Limit→3Agent Loop→

Want GPT or Codex?VibeHubGPT support

GPT-6 Astra

Build in 3D. Put your computer to work.