How Much Does It Cost to Run an AI Chatbot?

In model fees, a typical six-turn support chat costs between about $0.007 and $0.04 at current list prices, which puts 10,000 conversations a month somewhere between $70 and $420. The spread comes down to which model you pick and, less obviously, how long your conversations run. Here's how token billing works for chatbots, what a conversation costs on five popular models, and where the hidden costs come from.

How is AI chatbot usage billed?

Every major model API bills by the token, and a token is roughly three-quarters of an English word. You pay one rate for input tokens (everything you send the model) and a higher rate for output tokens (everything it writes back). Rates are quoted per million tokens.

Here are the list prices we used for the examples on this page. They change often, so check the LLM pricing tracker for today's numbers before you budget:

The part people miss is what counts as input. It's not just the customer's message. It's your system prompt (the instructions telling the bot how to behave), any documents you pull in for context, and, on every turn after the first, the whole conversation so far. Models don't remember anything between requests. Your app re-sends the history each time, and you pay for it each time.

How much does one conversation cost?

Take a typical support chat: a 1,000-token system prompt, customer messages of about 50 tokens, and bot replies of about 250 tokens. Over six back-and-forth turns, that works out to 10,800 input tokens and 1,500 output tokens. The input figure is so much larger because the prompt and history get re-sent on all six turns.

So a small business handling 10,000 chats a month is looking at somewhere between $70 and $420 in model costs, depending on the model. That's cheap next to a human agent. It's the scale and the conversation length that turn it into a real line item.

Why do long conversations cost so much more?

Because the re-sent history grows every turn, the total input grows faster than the number of turns. Double the length of a chat and you roughly triple the input bill. Same 1,000-token prompt, same message sizes:

A 20-turn chat costs about 5.6 times as much as a 6-turn one, for a bit over three times the turns. If your bot handles long troubleshooting sessions, that curve is your budget. Common fixes: summarize older turns instead of re-sending them word for word, cap how many turns the bot handles before offering a human, and keep retrieved documents short.

How much can prompt caching save?

A lot, for chatbots in particular. Your system prompt is identical on every request, and providers now let you cache repeated input. On current Claude and GPT-5 models, cached input reads bill at about a tenth of the normal input rate. (Anthropic also charges a small premium the first time a prompt is written to the cache, which washes out when the same prompt is hit thousands of times.)

Caching just the 1,000-token system prompt in the six-turn example drops 10,000 monthly chats from $366 to about $258 on Claude Sonnet 5, and from $285 to about $218 on GPT-5. That's a 24% to 30% saving without touching the prompt. Cache the growing conversation history too and it gets better, since that's where most of the input tokens live in longer chats.

What else ends up on the bill?

• Retrieval context. If the bot looks up help articles and pastes them into the prompt (RAG), a couple of 800-token articles per turn can double your input tokens. Trim to the relevant paragraphs.
• Reasoning tokens. Some models "think" before answering, and that thinking is billed as output. For simple FAQ bots, use the lowest reasoning or effort setting that still answers correctly.
• Retries and tool calls. A bot that checks order status via an API makes extra model calls per turn. Each one re-sends context.
• Hosting and platform fees. Model tokens are usually the smallest part if you buy a finished chatbot product. Per-seat or per-resolution pricing from a vendor can cost far more than the underlying API.

How do you estimate your own chatbot costs?

Work it out in four steps:

1. Paste your system prompt and a few real customer messages into the token counter to get actual token sizes instead of guesses.
2. Look at your chat logs (or your support tickets) for the average number of turns per conversation.
3. Total input = turns × (system prompt + message) + the re-sent history, which is (message + reply) × turns × (turns − 1) ÷ 2. Output = turns × reply length.
4. Multiply by your model's rates from the pricing tracker and by monthly conversation volume, then add 20% for retries and odd long chats.

Then test the cheaper model first. Budget-tier models handle a lot of routine support well, and the price gap is 5× or more. Move up only for the conversation types where the cheap model measurably fails. Well-structured prompts help small models more than big ones; the prompt library has templates worth starting from.

The bottom line

A typical six-turn support chat costs between about $0.007 and $0.04 in model fees at current list prices, so 10,000 chats a month runs roughly $70 to $420. Conversation length matters more than you'd expect, because history is re-sent every turn. Cache your system prompt, keep context lean, and start with a budget model.

Frequently Asked Questions

Related Tools