AI Agent Cost Calculator
Model what an agent actually costs: steps per run, sub-agent fan-out, context that grows as the transcript accumulates, cache hits, and retries. Gives cost per run, per developer and per month — and the date your budget runs out.
Where the money goes
| Re-sending the accumulated transcriptStep 19 pays again for everything steps 1-18 produced. This grows with the SQUARE of step count. | 42.9% |
| Sub-agent fan-out1 sub-agent x 8 steps, each running its own loop | 23.0% |
| Retries8.0% of steps retried — a retry pays full price and re-sends the context again | 6.9% |
| Cost per run | $0.798 |
These shares overlap and do not sum to 100%. Each is measured by switching that one factor off and taking the difference, because the factors compound — presenting them as a clean partition would be arithmetic theatre.
What this model assumes
- Context growth is quadratic, and that is the finding. Each step re-sends the system prompt plus everything previous steps added and produced. Over n steps that sums to roughly n²/2 times the per-step growth — which is why doubling an agent’s step count roughly quadruples its cost rather than doubling it.
- Caching is credited conservatively. Cached input is charged at the model’s published cache-read rate where the catalog has one, and at the full input rate where it does not. Cache WRITE premiums are not modelled, so heavy-caching estimates here are slightly optimistic in that one respect.
- Retries multiply steps, not just calls. A retried step re-sends the whole accumulated context, so retries are more expensive than their rate suggests.
- One model for the whole loop. Real agents often route cheap steps to a small model. If yours does, this is an upper bound — and the gap between this figure and your bill is the value of that routing.
- This is not the token calculator. The LLM Token Cost Calculator prices an API call and carries a flat agentic multiplier as a planning heuristic. This tool replaces that heuristic with the arithmetic. Use that one for a chat feature; use this one for anything that runs a loop.
Claude Sonnet 5 — $2.00 per 1M input, $10.00 per 1M output, $0.20 per 1M cached input. Verified 2026-08-04. Only models with a verified price are selectable.
Frequently asked questions
Why do AI agents cost so much more than a single API call?
Because each step re-sends the accumulated transcript. Step 12 pays again, as input, for everything steps 1 through 11 read and produced. Summed across a loop that is quadratic in step count, not linear — so doubling an agent’s steps roughly quadruples its cost. Add sub-agent fan-out and retries and a single "task" can cost tens of times what a token calculator would quote for the same prompt.
How much does it cost to run a coding agent per developer?
It depends almost entirely on loop length and cache hit rate, not on the model’s headline token price. A long-running coding agent with a large system prompt and heavy tool use can cost a few dollars per run; at a dozen runs a day that is tens of dollars per developer per day. Prompt caching is the single biggest lever — it commonly removes most of the input cost.
How do I reduce agent token costs?
In rough order of impact: raise the cache hit rate so the re-sent prefix is cheap; shorten the loop so the quadratic term is smaller; trim the system prompt and tool definitions, which are paid for on every single step; route cheap steps to a smaller model; and reduce retries, because a retried step re-sends the whole context again.
What is a realistic agentic cost multiplier?
There is no single number, which is the point of this tool. Reported overshoots against pre-agentic budget plans cluster around 2-3x, but that is an average across very different shapes. A five-step triage agent behaves nothing like a fifty-step coding agent with six sub-agents, and a flat multiplier hides exactly the difference that decides the budget.
How is this different from an LLM token cost calculator?
A token calculator prices one request: input tokens times a rate, plus output tokens times a rate. It is the right tool for a chat feature. This one models the loop — steps, fan-out, context accumulation, caching and retries — which is what actually determines the cost of anything autonomous. Both are on this site; use the token calculator for a single call.
A model is a forecast. Your invoice is the measurement.
Finitizer tracks real token consumption per agent, per feature and per team across Anthropic, OpenAI, Bedrock and Vertex — so you find out that a loop got longer from a trend line, not from a budget alert on the 22nd.