AI Agent Experiment Cost Calculator
Agent development is early. Every architecture, prompt, tool stack, memory system, and loop you test costs tokens. Model your experiment budget before the invoice models it for you. All math runs in your browser.
Model pricing
Illustrative sample prices. Edit to match current provider pricing.
| Model name | Input $/1M | Output $/1M |
|---|
No models yet. Add a model row to see cost estimates.
Loop simulator
Describe one agent run. Values below zero are clamped to zero.
Presets
Context growth per step
Input tokens the model reads at each step of a run.
Within context limitLive cost results
Monthly comparison
How this is calculated
Show formulas
history(i) = (i-1) * (avgOutput + contextGrowth) if sliding window: history(i) = min(history(i), windowSize) input(i) = systemTokens + history(i) inputTokens/run = sum over steps of input(i) outputTokens/run= steps * avgOutput cost/run = inputTokens * inPrice/1e6 + outputTokens * outPrice/1e6 cost/day = cost/run * runsPerExperiment * experimentsPerDay cost/month = cost/day * 30
FAQ
Why do agent loops get expensive?
Each step of a loop re-sends the entire conversation so far. A 20-step run does not cost 20 single calls; it costs 20 calls where each one carries all prior tool results and reasoning, so input tokens grow roughly quadratically with steps.
How does context growth multiply cost?
If each step adds tokens to history, step N reads everything from steps 1 through N-1. Doubling steps can quadruple input spend. Sliding windows or summarized memory cap that growth at a fixed window size.
A twelve-step loop re-reads earlier context
Read the explanation
This saved variant defaults to twelve steps, two thousand five hundred prompt tokens, three hundred fifty output per step and six hundred growth per step. Each added history step is nine hundred fifty. Repeated prompts total thirty thousand; zero through eleven sum sixty six times nine hundred fifty is sixty two thousand seven hundred history tokens. Input totals ninety two thousand seven hundred, with four thousand two hundred output tokens. At three pixels per thousand input tokens the prompt bar is ninety, history one hundred eighty eight point one and total two hundred seventy eight point one. This is a hypothetical local cost model, not actual model execution or tokenization evidence. Sliding mode caps the history component, not the system prompt. A hypothetical two thousand token window leaves first three histories zero, nine hundred fifty and nineteen hundred, and the remaining nine at two thousand. History total is twenty thousand eight hundred fifty; add thirty thousand prompts to get fifty thousand eight hundred fifty input. Compared with full history ninety two thousand seven hundred the saved amount is forty one thousand eight hundred fifty. At three pixels per thousand the bars are two hundred seventy eight point one, one hundred fifty two point five five and one hundred twenty five point five five. The red context line only warns; exceeding it does not halt the run or change the token sum. At the illustrative three dollar input and fifteen dollar output prices per million, ninety two thousand seven hundred input costs twenty seven point eight one cents and forty two hundred output six point three cents. Total thirty four point one one cents. At eight hundred pixels per dollar input measures two hundred twenty two point four eight, output fifty point four and total two hundred seventy two point eight eight. Twenty five runs times six daily experiments gives one hundred fifty runs and fifty one dollars sixteen and a half cents daily, or fifteen hundred thirty four dollars ninety five monthly. The display rounds high prices to whole dollars. Native presets, model fields, history mode, reset and inline SVG charts are local. These examples cannot establish current provider prices, real agent accuracy or billed savings.