How Much Does It Cost to Run an AI Chatbot?
The honest answer is a range so wide it surprises most founders: the same chatbot, serving the same users and answering the same questions, runs $6.00 a month on the cheapest model we price and $1,080 on the most expensive. The cost formula is simple: users × requests per user × (input tokens × input rate + output tokens × output rate). Everything else is picking the numbers honestly.
A realistic example, priced three ways
Work a realistic example: a customer support bot for a product with 5,000 monthly users, each triggering about 8 requests a month, with a 1,200 token prompt (system instructions plus knowledge base context plus the customer's message) and 300 token replies. That is 40,000 calls, 48 million input tokens, and 12 million output tokens per month. On Ministral 3 3B 2512, the cheapest model in our basket, that workload runs $6.00 a month. Half the models we price run the same bot for between $25.05 and $180.00. On GPT 6 astra, the most expensive we price, the identical bot costs $1,080. Same bot, same users, same questions, and a 180x spread decided entirely by which model answers.
Four levers that close the gap
Model tier. Most support conversations are routine, and the spread above is decided almost entirely by this one choice. It is also the only lever that can change your answers, so it is the one to test rather than assume.
Prompt caching. That 1,200 token prefix is mostly identical between calls, and where a cache-read rate is published the median discount is 90 percent off standard input. Note the ceiling though: only the fixed part qualifies, so on this workload the saving applies to the instructions and knowledge base, never to the customer's own message.
History discipline. Resending the whole conversation on every turn is the most expensive default in chat, and it compounds: a ten-turn conversation can cost several times a two-turn one for the same number of answers. Summarize old turns or drop them.
Down-routing. Answer the easy majority with a cheap model and escalate only what fails a quality check. This is the most work to build and it is what turns the spread above into an actual bill somewhere near the bottom of it.
What the formula leaves out
The arithmetic above is honest about what it covers and it does not cover everything, so add margin for four things before you quote anyone.
Retries and failures. Timeouts, refusals and malformed outputs get retried, and a retry is billed like any other call. Nobody knows their retry rate before building, which is exactly why it belongs in the margin rather than the estimate. Your own traffic. Evaluation runs, test conversations and demos all bill at the same rate as customers. The first version is talkative. Early prompts are longer and early answers are wordier than the ones you ship three months later. Reasoning tokens, on models that produce them, bill as output and are not predictable in advance.
The number that actually matters is cost per conversation
Monthly totals are what founders ask for and the wrong unit for deciding anything. Divide by conversations instead, and the question becomes answerable: at 40,000 calls a month the example above works out to fractions of a cent per exchange on a budget model and a few cents on a flagship.
That is the figure to hold against what a conversation is worth to you. If a support conversation deflects a ticket that would have cost a person fifteen minutes, almost every model on the board is defensible and you should optimise for answer quality. If it is a free-tier feature with no revenue attached, the cheap end is the only end that makes sense. Costs scale close to linearly with use, so ten times the users costs roughly ten times as much, with no seat licence and no step change. That property is what makes the per-conversation figure worth computing once and trusting.
The Customer Support Bot preset in the advanced calculator loads this exact scenario; change the sliders to match your product and the ranking updates instantly.