The economics of agent tokens: Calculating cost per resolved business task

Chief Financial Officers and founders frequently panic when they review monthly OpenAI or Anthropic invoices and see volatile, unpredictable usage fees. But looking at raw token expenses in isolation is like analyzing a logistics fleet by looking only at the fuel bill without counting the packages delivered.

The correct financial metric for autonomous systems is Cost per Resolved Task (CPRT). CPRT ties token expenditure directly to commercial business value.

Calculating Cost Per Resolved Task (CPRT)

CPRT is calculated by summing the total compute and token expenditures across all model tiers and dividing by verified successful business outcomes:

• Input and Output Token Fees: Aggregated across reasoning, tool calls, and final generation passes.
• Infrastructure Overhead: Serverless execution, vector database queries, and caching layer fees.
• Failed / Retried Execution Costs: Factoring in the token burn of aborted attempts and error states.
• Task Resolution Verification: Only counting tasks that met quality thresholds without requiring manual human remediation.

A $2.00 token cost is a tragedy if it failed to answer a simple FAQ; it is a miracle if it successfully onboarded a $50,000 client.

Optimization strategies

Once you measure CPRT, you can optimize ruthlessly: route simple reasoning steps to lightweight, low-cost models, cache recurring semantic queries, and reserve frontier reasoning models only for high-value decision gates.