Free tool

Agent Burn Calculator

Steady-state agent spend is a budgeting problem. The number that ruins a quarter is the loop that did not terminate at three in the morning. This works out both.

Workload

Completed user-facing tasks, not individual model calls.

Median across your traces. Guess high if you do not know.

%

Extra calls from failed tools, reflection loops, and re-planning.

Per call

Including the system prompt and every re-injected turn.

$/M tok
$/M tok

Steady state, per day

$414

18,400 tool calls at $0.02 each.

Steady state, per month

$12,602

This is the number finance already knows about.

When a loop does not terminate

Tasks that enter an unbounded retry or reflection loop.

A tight loop against a fast tool clears several hundred an hour.

Without a spend alert, this is 'until the invoice arrives'.

calls / task

Hard cap per task at the tool-call layer, enforced at runtime.

Worst case, per day

$779

16,200 unbudgeted calls on top of normal traffic — 1.9× your steady state.

Worst case, per month

$23,698

If the loop is not a one-off and nothing changes.

Same day, with the breaker

$417

Each looping task stops at 40 calls no matter how long detection takes.

Avoided, per month

$11,013

The difference between finding out in milliseconds and finding out on the invoice.

Runs entirely in your browser. Nothing you type here is transmitted, logged, or stored anywhere. There is no server-side copy because there is no transmission. Close the tab and it is gone.

How this works, and what it does not cover

Cost per call is input tokens times the input rate plus output tokens times the output rate. Daily spend is tasks multiplied by tool calls per task, plus your retry overhead, multiplied by cost per call. Monthly figures use 30.44 days.

The runaway half assumes a looping task keeps calling at its natural rate until something stops it. A circuit breaker at the tool-call layer stops it at a fixed number of calls, which is why the capped figure does not move when you change the detection time. That independence from detection latency is the entire argument for enforcing at the tool layer rather than alerting on a dashboard.

What this does not model: prompt caching, batch discounts, provider commitments, or the human cost of the incident. Every one of those makes the real gap between the two columns wider, not narrower.

The published $40K/day figure came out of this arithmetic on a real production system — the incident write-up is here.

The written teardown: the seven integration patterns that account for most of the spend in the codebases behind the $200K/yr figure, and the arithmetic for each.

No cookies, no tracking pixels, never shared or sold. How the data is handled.

If you would rather not do this yourself

LLM Cost & Routing Sprint

Three weeks: find where the money goes, fix the largest lines, and leave the controls that stop it coming back.

3 weeks · $15-20K
All free tools