AI Agent Cost Calculator

How the estimate is calculated

This calculator is an illustration, not a benchmark. It runs entirely in your browser and sends nothing anywhere. It multiplies four inputs you control by a list-style token price for the model tier you pick:

• Calls per month = number of agents × LLM calls per agent per day × 30.
• Cost per call = input tokens ÷ 1,000 × the input price, plus output tokens ÷ 1,000 × the output price.
• Monthly spend = calls per month × cost per call. Annual spend is the monthly figure × 12.
• One runaway agent assumes a single agent loops at 20 times its normal daily call volume for a month. The figure shown is the extra cost on top of the fleet's normal spend.

The three tiers use rounded prices per million tokens: frontier (large) $5 input and $15 output, mid-tier $1 and $4, small or fast $0.20 and $0.80. They stand in for the shape of current provider price lists, not for any one vendor's rate card.

What the estimate leaves out

Real agent spend rarely follows one average. The model does not include prompt caching or batch discounts, retries after errors or timeouts, embeddings and reranking, tool-call round trips that resend context, growing context windows in long sessions, or negotiated enterprise pricing. Use your provider invoice and per-agent usage data to replace these inputs before you plan a budget on them.

Why the runaway figure matters

Agent cost grows with fleet size and call volume, and in this model one agent looping at 20 times its volume costs as much as 19 extra agents running normally. A billing alert reports that spend after it happens. A hard cap checked before the call is dispatched stops it. The runaway agent spend threat model walks through how these loops start, and budget policies in Praesidia (Team and Enterprise plans) shows how new agent tasks and over-budget gateway model calls are refused once a threshold is reached. For forecasting from measured usage rather than averages, see token usage forecasting.