The Cost Components
Every agent task has three cost components: input tokens (context the model reads), output tokens (text and tool calls the model generates), and tool execution costs (API calls, compute, etc. from the tools themselves).
Input tokens are the biggest cost driver because context grows with every step. Step 1 might have 3,000 tokens of context. By step 10, the context includes all previous tool results and reasoning, which might be 50,000 tokens. The model pays for all of that context at every step.
A Quick Estimation Method
Estimate the number of steps: How many tool calls will the agent need? For a simple lookup, maybe 2-3. For research with synthesis, maybe 10-15. For complex multi-source analysis, maybe 20-30.
Estimate context growth per step: Each tool result adds 1,000-5,000 tokens typically. A database query result might be 2,000 tokens. A file read might be 5,000. A web search result might be 3,000.
Calculate the total: For N steps with average context growth of G tokens, total input tokens are roughly: N * (initial_context + N*G/2). The "/2" accounts for context growing linearly, so the average step processes about half the final context.
For a 10-step task with 3,000-token tool results: 10 * (3000 + 10*3000/2) = 10 * 18,000 = 180,000 input tokens. At $3 per million input tokens, that's about $0.54. Add output tokens (maybe 20% of input cost) and you're looking at roughly $0.65.
When Costs Spiral
Costs spiral when agents get stuck in loops, when tool results are unexpectedly large, or when the task turns out to be more complex than anticipated. Setting a maximum step count and a maximum token budget on your agent prevents runaway costs. Most agent frameworks support both limits.
Caching also helps significantly. If the agent queries the same data twice, serving the second query from cache avoids both the tool cost and the token cost of processing the duplicate result. For agents that run similar tasks repeatedly, caching can reduce costs by 30-50%.
Related Reading
- The Cost Economics of Running AI Agents
- Why Most AI Agents Fail at Multi-Step Reasoning
- Why Context Window Size Matters for Tool-Using AI
Discover AI agents on Skillful.sh. Search 137,000+ AI tools.