AI chatbot development price: what drives the cost

What you're really asking when you ask for a price

When a founder asks us for an AI chatbot development price, they're usually bundling three separate questions into one: what will it cost to build, what will it cost to run each month once it's live, and how much of that number holds up once real users start typing weird things into the box. Those three numbers don't move together. A chatbot that costs 8,000 euros to build can still cost more to run in year two than it cost to make, if nobody watched the token bill. None of that is a knock on quotes that only cover the build. It's just worth knowing which question a number actually answers before you compare two of them.

That's the part most quotes skip. The build is a fixed project cost. The model calls behind it are a variable cost that scales with usage, and it can grow or shrink based on decisions nobody thinks to ask about at kickoff: which model, how much context you feed it, and whether you're willing to share your usage data for a lower rate.

The three things that make up an AI chatbot development price

Three line items decide the number on your invoice, and only one of them is negotiable with a developer.

The first is the build itself: the hours to design the conversation flow, wire it into your CRM or booking system, write and test the prompts, and handle the cases where the model gets it wrong. This is the part an agency quotes as a fixed fee or a day rate, and it's the part most comparison shopping focuses on. The build hours don't stay constant either. A chatbot that answers FAQs from a fixed script is a different job from one that checks your booking calendar, updates a customer record, and hands off to a human when it gets stuck. The second kind touches more of your existing systems, which is where the hours actually go, not in getting the model to sound polite.

The second is the model itself, billed per token by whoever makes it. Anthropic's newest Claude Fable 5.1, released this week, prices input tokens at $10 per million, output at $50 per million, and cache writes at $12.5 per million. Say your chatbot handles 10,000 conversations a month, each one running to roughly 1,000 input tokens and 300 output tokens: that's 10 million input tokens and 3 million output tokens, which puts the model bill for that month at around $250, before anyone has paid a developer for a single hour of work.

The third is maintenance: someone watching what the bot says when a customer asks something the prompt didn't anticipate, and adjusting it. This is the line item that gets forgotten in a first quote and shows up as a surprise three months in.

Why a price cut on paper does not mean a lower bill

Anthropic cut the price of reading from its prompt cache by 75%, from a dollar down to $0.25 per million tokens, with Fable 5.1. On paper that looks like a straightforward saving for anyone running long conversations through the API, which is exactly what a chatbot does all day.

In practice, Artificial Analysis measured the new model producing about 1.7 times more output tokens per task than its predecessor. Output tokens are priced far higher than cached reads, so the net effect for a typical task was a 20% increase in cost, not a cut. A single number in a pricing announcement rarely tells you what your bill will actually do. What matters is how many tokens your specific chatbot burns through per conversation, and that depends on how it's built, not on the sticker price of the model behind it.

This is one of the reasons we test a chatbot's actual usage pattern against two or three models before we lock one in, rather than picking whichever announcement is loudest that month.

The data-sharing discount, and why we'd think twice before taking it

Meta's new Muse Spark model, aimed at coding and other agent work, comes with a second pricing tier most people miss on the way to checkout. Its standard rate is $1.25 per million input tokens and $4.25 per million output tokens. Agree to let Meta use your prompts and outputs to train future models, and those same tokens cost $0.10 and $0.20, a discount averaging around 95%.

For a side project, that trade is easy. For a business chatbot that handles customer questions, order details, or anything with a name and an email address attached, it's a different conversation. Arvind Narayanan, a computer science professor at Princeton, has pointed out that large companies already vote with their wallets on this: they stick with token-billed enterprise plans instead of subscription plans priced 10 to 20 times cheaper, because the subscription tiers train on their data and the enterprise ones don't. If a discount like Meta's ever gets folded into a chatbot build, it belongs in the contract in plain language, not buried in a rate card.

How we'd actually price a project

We charge by the work: scoping the conversation, building the integration, testing it against real questions your customers actually ask, and a maintenance retainer once it's live. We don't build the token markup into a mystery number, because you should be able to see what you're paying the model provider versus what you're paying us.

What we'd skip, and tell a client to skip: locking into the newest model the week it launches because the benchmark chart looks good. Fable 5.1 has real strengths for coding and long multi-step tasks, but a customer-facing chatbot answering "where's my order" doesn't need the most expensive frontier model on the market. Cheaper models handle narrow, well-scoped conversations fine, and the savings compound every month the bot is live.

If you're at the stage of comparing quotes and the numbers don't line up, that's usually because one of them includes ongoing model costs and the other doesn't. Ask directly. It's the single question that explains most of the gap between a 5,000 euro quote and a 20,000 euro one for what sounds like the same chatbot. We cover this kind of scoping question directly as part of our AI services, including the part where we tell you if a smaller model or a simpler flow will do the job for less.

One more thing worth checking before you sign anything: whether the vendor's price includes watching the bot's real conversations for the first month and fixing what breaks. A chatbot that works in the demo and drifts once real customers start typing typos, slang, and half-finished sentences is the most common way a fixed-price quote turns into a second invoice.