LLM Cost Calculator
for AI Agents.

How to Read Your Results

The calculator produces five outputs. Four are straightforward arithmetic — the one that requires explanation is the loop risk multiplier. Here is what each setting means in practice and when to use it.

The normal daily cost and normal monthly cost outputs assume no loops occur. They represent your expected operating cost under normal conditions. If your actual spend is consistently higher than the normal daily estimate, you may already have a low-level loop running — an agent that is retrying a failing tool call 2-3× per session instead of once.

What "Recommended Daily Cap" Means

The recommended daily cap is your calculated normal daily cost multiplied by 1.2 — a 20% safety buffer. This is not an arbitrary number. It reflects two practical realities of production agent traffic.

First, legitimate traffic spikes are real. If your agent normally handles 50 sessions per day, it may handle 55 on a busy day, or 60 when a marketing campaign drives extra traffic. A cap set at exactly your normal daily cost would trip the circuit breaker on those legitimate spikes, causing false positive halts and degrading your service. A 20% buffer absorbs normal variance without letting runaway loops through.

Second, the cost of a false positive (agent halts unnecessarily) is much lower than the cost of a false negative (loop runs unchecked). At a 20% buffer, you have zero false positives from normal traffic variance in the vast majority of production deployments. And any cost that exceeds 120% of your baseline is, by definition, anomalous — worth investigating whether it is legitimate or a loop.

For high-variance workloads — agents with unpredictable session counts, or agents that handle large documents with variable token counts — consider increasing the buffer to 1.5× or 2×. For agents with highly predictable traffic, 1.1× is sufficient.

How to Implement Your Cap

Take the recommended daily cap value from the calculator above and plug it directly into the SupraWall protect() wrapper. The configuration below uses your current calculated values and will be enforced at the SDK level — no agent can bypass it regardless of its prompt instructions.

from suprawall import protect

secured = protect(agent, budget={
    "daily_limit_usd": 72.00,    # your recommended cap
    "circuit_breaker": {
        "strategy": "combined",
        "max_identical_calls": 10,
        "window_seconds": 60,
    },
    "on_budget_exceeded": "halt",
})

The daily_limit_usd field sets a hard dollar cap per agent per day. When the cap is reached, SupraWall raises a BudgetExceeded error and blocks all further tool calls until midnight UTC. The circuit breaker configuration provides a second layer of protection: it will halt the agent if it detects loop patterns before the dollar cap is hit. See the Budget Limits documentation for the full configuration reference.