How much does an AI agent actually cost to run?
The honest cost model for an AI agent: inference, tools, data, and the runaway risk that the sticker price never shows. Plus the numbers to budget.
Everyone quotes per-token prices. Nobody quotes what an agent costs to run, because the real cost is rate × volume × behavior — and behavior is the variable. Here's the model we use when talking to teams.
The three cost lines
Inference. The model bill: per-token rates times the tokens an agent actually consumes. Long-context runs and retries inflate this line faster than any rate card suggests — see the context-cost math.
Tools and data. Every API the agent calls: search, enrichment, compute, data vendors. This is the line most teams don't see until the invoice — and the one allowlists govern.
The runaway risk. Not a line item — a multiplier. A retry loop turns one failed call into 40. The incident database's runaway costs show the range: hundreds to millions.
The number to budget
Work backwards from the task: estimate legitimate daily spend, multiply by 1.5, and make that the daily ceiling. Then let the audit log tell you the real number. The price of the agent isn't the token rate — it's the ceiling you enforce.