The short answer
An agent run is not one model call. It is a loop of calls, and every call pays for the tokens it reads and writes. The cost of one run is:
cost per run = steps × (1 + retry rate) × cost per step
cost per step = (fresh input tokens × input price
+ cached input tokens × cache-read price
+ output tokens × output price) ÷ 1,000,000Prices are quoted per million tokens. Multiply the cost per run by runs per day to get the daily bill, and by days per month for the monthly one.
What drives the number
- Input tokens per step. Agents resend their context (system prompt, tool definitions, history, files) on every step, so input usually dominates.
- Steps. A coding agent can take dozens of tool calls per task; a triage bot may need two.
- Retries. Failed tool calls and invalid outputs repeat a step. A 10% retry rate means 1.1× the calls.
- Prompt caching. The repeated part of the context can be billed at the cheaper cache-read price, which often cuts the input bill sharply.
- Model choice. Routing simple steps to a cheaper model of the same vendor lowers the average step price.
Typical agent profiles
These are the starting presets in the AgentTab calculator (assumptions, not measurements):
| Agent | Input tokens/step | Output tokens/step | Steps | Retry rate | Cacheable |
|---|---|---|---|---|---|
| Coding agent | 40,000 | 2,000 | 25 | 10% | 80% |
| Research agent | 60,000 | 3,000 | 15 | 8% | 60% |
| Customer support bot | 6,000 | 400 | 4 | 5% | 70% |
| Inbox triage bot | 3,000 | 300 | 2 | 3% | 50% |
For a coding agent that means 25 × 1.1 = 27.5 model calls per run, each reading 40,000 tokens: about 1.1 million input tokens per run before caching.
Get your own number
Open the cost calculator — current list prices for many models, with caching and routing savings itemized. Want to know whether a pricier agent is worth it? Compare scores in the Arena.