AI agent step-cost visualizer
Each agent step re-sends the whole context. Tool results pile up — cost grows faster than step count. Tune the assumptions and see the per-step bill.
Assumptions
Totals
Per run
Per day
Per month (30d)
Input tokens / run
Cost per step
Input billed = initial context + (step − 1) × tokens added. Each row is one LLM call in the agent loop.
| Step | Input tokens | Output tokens | Step cost | Cumulative | Relative |
|---|---|---|---|---|---|
Pricing snapshot from July 2026 vendor pages. No prompt caching modeled — real agents with cache hits pay less on repeated prefix tokens.
About this tool
Agent loops are sneaky expensive. Every step sends the full conversation plus tool results back to the model. If step 1 bills 4k input tokens and each step adds 1.5k of tool output, step 8 bills ~14.5k input — and you pay for all of it again.
That is why agent cost scales worse than linear: more steps mean longer context, and longer context means every subsequent step costs more. The fix is not always a cheaper model — it is summarizing tool results, trimming history, and caching stable prefixes.
This calculator uses simple arithmetic, not your real trace. Use it to sanity-check whether a 12-step ReAct loop at 100 runs/day is a rounding error or a line item.
Read more
- Prompt caching, compression, and compaction
- Agency agents — multi-step AI workflows
- Executor — running agentic tasks
- Sample app ideas — projects to build and cost out
- AI and the future of software developers
Free books, courses, and software
Get my programming books, course editions, and complete software.
Get the download library →