How Much Does an AI Chatbot Cost?

"How much does an AI chatbot cost?" has two answers: what you pay your provider, and what it actually costs to run. Understanding the second helps you read the first, and spot pricing that could hand you a nasty surprise after a busy week.

What actually drives cost

Almost all of the running cost of a support chatbot comes from the language model generating answers. That cost is driven by tokens (roughly, the words going in and out) and by which model does the work. Three levers dominate:

• Conversation volume: how many visitor sessions you handle in a month.
• Model choice: a frontier model can cost many times more per answer than a small, fast one.
• Tokens per answer: how much context is fed in and how long the responses are.

Embeddings, the indexing step that powers retrieval, are cheap by comparison, even for a large knowledge base, and only run when your content changes. The expensive part is generation, every time someone chats.

Why volume-based pricing exists

Because cost scales with usage, most providers meter on message credits: one credit per AI reply, sometimes weighted by which model answers. Pricing by reply lines the bill up with the thing that actually costs money, rather than charging a flat fee that is either a rip-off at low volume or a loss-maker at high volume.

The hidden risk: usage-based bills

Pure usage-based pricing has a sharp edge. If your chatbot goes viral, gets embedded on a high-traffic page, or is hit by an abusive script, an uncapped bill can balloon overnight. The cost is real on the provider's side, and someone has to pay it. The question is whether that someone is surprised.

Guardrails that prevent surprise bills

This is why we price in prepaid credits and publish the numbers. A few mechanisms keep a traffic spike from becoming an invoice shock:

• Prepaid credits mean the assistant can never bill past what you have already bought.
• Optional auto-recharge with a monthly dollar limit you set yourself, so topping up is your choice, on your terms.
• Per-visitor rate limits on the public widget to blunt abuse.
• A semantic cache that reuses answers to repeated questions instead of paying the model again.

Together these mean a viral day raises your traffic, not a bill you never agreed to. You can read exactly what each plan includes on the pricing page. The numbers are published on purpose.

Questions to ask any provider

Sticker price tells you little on its own. Before you commit, get clear answers to a few questions that separate predictable pricing from a meter that can run away from you:

• What exactly counts as a billable unit, and when does it reset?
• Is spending prepaid and capped by me, or can usage keep billing past my plan?
• Does a cheaper plan quietly downgrade answer quality when usage climbs?
• Are there per-visitor rate limits on the public widget to stop abuse?
• What happens when I run out, and does the assistant stop, throttle, or auto-charge on my terms?

A provider that answers these plainly is one you can budget around. Vague answers are a sign the cost risk has been left on your side of the table.

So what should you budget?

For most small and growing sites, a plan in the low tens of dollars a month covers serious volume (thousands of answered questions) as long as the assistant runs efficiently and the guardrails above are in place. The number to watch is not the sticker price but the worst case: a provider whose pricing is prepaid and published is telling you what your bad month looks like. One that is not is asking you to find out.