AI Cost Calculator

Most teams ship AI features without ever drawing a line from "tokens per request" to "annual invoice." Then the invoice arrives and someone has to explain why a chatbot costs $40,000 a year, or why moving from Claude 3.5 Sonnet to Haiku quietly saved $28,000 with no quality drop on the actual workload. This calculator closes that gap. You give it three honest numbers — average input tokens per task, average output tokens per task, tasks per day — and a provider/model choice. It returns cost per task, cost per month, and cost per year using realistic published prices as of June 2026. The math is deliberately boring: input tokens times input price plus output tokens times output price, divided by one million, multiplied by run rate. No volume discounts assumed. No batch API discounts assumed. No caching credits assumed. Those things are real and they help — and they are out of scope here because they vary by contract, vary by month, and tempt people to be optimistic about their own usage. This is the upper-bound estimate. The number your CFO should believe before you negotiate it down. A few things the calculator does NOT include, and which you should think about separately. It does not model retries, tool calls, agent loops where one user task fans out into ten model calls, or systems where Claude generates a draft and GPT critiques it. Those multiply real cost by 3x to 10x. It does not include embedding costs, vector store costs, or the human review labor that often dominates the bill. It assumes "local-ollama" is free at the API line — but you are paying for the GPU, the electricity, and the engineer who keeps it running, so call it $0.0001/M tokens of opportunity cost and decide for yourself. Use this as a sanity check before you commit to a model. Run it on three workload sizes (today, 6 months, end of year) and look at the year-end number. If that number scares you, you have a model selection problem, not a usage problem.