METHODOLOGY DEEP-DIVE · VERIFIED 15 JUL 2026
Why AI agents cost 10-15x more than you'd guess
Every calculation on this site so far has treated a "call" as one input, one output. Agentic workflows — an AI that plans, calls tools, reads the results, and iterates — break that assumption completely. A single user-facing "agent task" is actually several API calls stacked together, and the context sent on each one keeps growing. Estimate based on the final response alone, and you'll badly underestimate the real bill.
What actually happens inside one "agent task"
A typical multi-step agent task looks something like: plan the approach, call a tool, read the tool's output, decide on the next step, call another tool, and finally synthesize an answer. Each of those steps is a separate API call, and critically, each one resends the accumulated conversation history — the original request, every prior tool call, every intermediate result — because the model needs that context to make its next decision.
| Step | Input tokens | Output tokens |
|---|---|---|
| 1. Initial plan | 1,000 | 300 |
| 2. Tool call 1 + read result | 3,000 | 300 |
| 3. Tool call 2 + read result | 6,000 | 300 |
| 4. Tool call 3 + read result | 10,000 | 300 |
| 5. Reasoning / decision | 15,000 | 300 |
| 6. Final synthesis | 21,000 | 300 |
| Total for one task | 56,000 | 1,800 |
The gap, at 10,000 tasks/month
That's roughly a 15x gap between what a naive "cost per response" estimate suggests and what the workload actually costs once every intermediate step is counted honestly. On a pricier model tier, the absolute gap is larger even though the ratio holds — the same 6-step task on Claude Sonnet 5 runs closer to $1,300/month at that volume, not the few hundred dollars a final-response-only estimate would suggest.
Why this surprises people specifically with agents
A simple chatbot has one input, one output, and the math is intuitive. An agent's context grows with every step because each step needs to "remember" what came before — there's no way around resending that history unless the architecture specifically avoids it. The more tool calls or reasoning steps a task requires, the steeper that growth curve gets, which is exactly why agentic workloads are the category most commonly underestimated in cost projections.
The two things that actually control agent cost
- Number of steps. Every additional tool call or reasoning step doesn't just add its own tokens — it adds the entire accumulated history again on the next call. Fewer, more decisive steps compound savings faster than a cheaper model choice does.
- Prompt caching on the repeated prefix. Since most of what's resent on each step is identical to the previous call plus one new addition, this is close to a best-case scenario for prompt caching — see our caching guide for how much that can realistically claw back.
How we'd actually approach this
Before optimizing model choice for an agentic workload, count your actual steps per task and estimate the real accumulated context, not the final response size. Then check whether prompt caching applies to your architecture — it usually does for agent loops specifically, since the repeated-prefix pattern is almost exactly what caching is built for.
Worked example is illustrative — step count and context growth vary significantly by agent architecture and task complexity. Use the calculator with your own total accumulated tokens per task for an accurate estimate.