Practical guide
AI Agent Pricing Models: Cost per Run, Task, or Outcome?
Choose an AI agent pricing unit by tracing model steps, tool charges, retries, and successful outcomes before setting a budget.
Updated · Sources checked
Pick the unit your product actually sells
Agent costs can be reported per model call, run, completed task, or business outcome. Calls are easy to meter but poor for planning because a run may branch or retry. Cost per completed task is usually more useful: it combines successful and failed run spend over the number of accepted results.
Take an invoice-processing agent that handles 2,000 documents. A straight-through document averages four model calls; 15% require a second eight-call review path. The weighted average is 4 + (0.15 × 8) = 5.2 calls per document. Enter 5.2 only after converting the workload to a whole-number scenario—for example, 52 calls representing ten documents—with average tokens and apportioned tool cost in the AI Agent Cost Calculator. Divide that result by ten for cost per attempted document, then adjust for the final acceptance rate.
Decide what belongs in the price
Include paid browser, search, database, or external API calls through the tool-cost input. Keep human review, orchestration hosting, storage, monitoring, and support in a separate operating-cost line because the calculator does not model them. If a vendor quotes per resolution, clarify whether escalations, failed attempts, and repeat conversations are billable.
The tool uses GPT-5 rates and assumes the same average token and tool charge for every call. It does not compare agent platforms or price reasoning meters, caching, storage, or discounts. Use trace data from a pilot before committing to a customer-facing per-task price.
FAQ
Which pricing model is safest for an early agent?
A metered internal budget per attempted task is easiest to audit. Outcome pricing needs a precise outcome definition and enough history to estimate failure and escalation rates.
When should two agent workflows be modeled separately?
Separate them when their call graphs or paid tools differ materially, such as routine answers versus cases that trigger browsing and human approval.