Building a business chatbot? How to estimate usage costs.

By Ordinal Science · 2026-10-09

If you are building a chatbot for your business, one of the first questions is which model should power it. Most people start with ChatGPT from OpenAI or Claude from Anthropic. Others look at Google's Gemini, or at open models like Meta's Llama or Mistral that you can run yourself.

Cost is often the deciding factor, and it is also the easiest one to get wrong. The prices on a provider's website tell you very little until you know how your chatbot will use them. Here is how we suggest estimating what each option will really cost.

Start with the job, not the brand.

Before comparing prices, write down what the chatbot is for. Is it answering customer questions from your help articles? Summarising long documents for staff? Drafting emails? Pulling numbers from reports?

Each of these tasks uses a different amount of text. A customer support bot sends short questions but may need help articles with every answer. A document assistant reads long files, so input can be very large. A drafting tool writes a lot, so output matters more.

Be as specific as you can. "Help customers" is too vague. "Answer questions about delivery times, returns and product sizes, using our 200 help articles, and pass billing disputes to a person" is much better. A description like that tells you how much text goes in and out, which is what drives the bill.

Once you know the job, you can estimate costs against it. Without that, any figure you come up with is a guess.

Decide how you will pay.

There are two main ways to pay for an AI assistant.

The first is a per-person subscription. ChatGPT and Claude both offer team or business plans for staff, charged per user per month. Check whether other providers, such as Google, offer something similar for your setup. This suits internal assistants that people use directly. The cost is simple to work out: number of users times the monthly fee.

The second is paying for usage through the provider's API. This is how most customer-facing chatbots work. The cost depends on how much text goes through the model, so it takes more effort to estimate. The rest of this article focuses on that kind of pricing.

Understand how usage pricing works.

Most providers charge by the "token." A token is a small chunk of text, often part of a word. As a rough guide, 1,000 tokens is about 750 English words. That rule of thumb comes from OpenAI, and other providers count tokens differently. One measurement put Claude at roughly 550 to 650 words per 1,000 tokens. So the same text can use more tokens, and cost more, on one model than another. You pay for the text you send in (the input) and the text the model sends back (the output). Output tokens usually cost more than input tokens, typically five to eight times more.

Each provider also offers several model sizes. OpenAI has its full GPT models and smaller "mini" versions. Anthropic has Claude Haiku (small and fast), Sonnet (mid-range) and Opus (more capable), and its line-up has grown beyond these three over time. Google has Gemini Flash (fast and cheap) and Gemini Pro (more capable). The larger models cost more per token. The smaller ones can be ten times cheaper or more. Prices and line-ups change often, so check the current rates on each provider's website rather than relying on old comparisons.

To estimate your costs, work through a simple calculation. First, estimate how many conversations you expect per month. Second, estimate the average length of a conversation, including any background documents you feed in. Third, multiply by the per-token price for input and output separately.

That last point about background documents matters. If your bot looks up company information and passes it to the model with every question, your input costs can be much higher than the questions alone suggest.

Conversation history matters too. Most chatbots resend the whole conversation with each new message so the model remembers what was said. A five-message chat does not cost five times one message. It costs more, because each message carries everything before it.

A worked example. (Prices checked in November, 2026.)

Here is a simple example for a customer support bot. The prices below are hypothetical round numbers for illustration only. They are based on list prices some providers charged in 2024 and 2025. Some current rates are lower, so plug in today's prices from each provider's pricing page.

- Large model: $3 per million input tokens, $15 per million output tokens. - Small model: $0.25 per million input tokens, $1.25 per million output tokens.

Say the bot handles 10,000 conversations a month. Each conversation has a few back-and-forth messages. With every message, the bot sends the model the customer's question, the conversation so far and a few relevant help articles. That adds up to about 15,000 input tokens per conversation. The bot's replies add up to about 1,000 output tokens per conversation.

So each month you send 150 million input tokens and receive 10 million output tokens.

On the large model: - Input: 150 × $3 = $450 - Output: 10 × $15 = $150 - Total: $600 a month

On the small model: - Input: 150 × $0.25 = $37.50 - Output: 10 × $1.25 = $12.50 - Total: $50 a month

Notice where the money goes. Most of the cost is input, and most of the input is the help articles. If you only sent the questions, input might be closer to 1,000 tokens per conversation. That would cut the large model bill from $600 to about $180. Sending fewer, better-chosen articles is often the easiest saving you can make.

Some providers also offer discounts for "caching." This means text you send repeatedly, like your standard instructions, is billed at a lower rate. It is worth asking about if your bot reuses the same background text often.

Some also offer a lower rate for "batch" jobs. These are tasks that don't need an instant answer, like summarising yesterday's support tickets overnight. If part of your workload can wait, it can be cheaper.

Plan for a range, not a single number.

Your first estimate will be wrong in some way. Usage may be higher than expected, or conversations may run longer. So build three versions: a low case, an expected case and a high case. For example, try half your expected volume, your expected volume and double it.

If the high case is still affordable, you are in good shape. If it isn't, you know to set limits or plan a cheaper setup before launch.

Measure real usage early.

Guesses about token counts are just guesses. The quickest fix is a small pilot. Run the chatbot with a few real users or a batch of real questions for a week or two. Every major provider shows token usage in its dashboard. Use those numbers to redo your estimate.

Most providers also offer free tools that count the tokens in a piece of text. Paste in a typical question, your instructions and a sample help article. You will get a much better figure than a word count alone.

Don't pay for more power than you need.

It is tempting to use the most advanced model for everything. Often that is wasteful. Many routine tasks, like sorting messages or answering simple questions, work well on a smaller model.

A common approach is to mix models. Use a cheap, fast model for easy requests and send only the harder ones to a more capable model. This can cut costs a lot without hurting quality.

Using the example above, suppose 80% of conversations are simple and go to the small model. The other 20% go to the large model.

- Small model, 8,000 conversations: about $40 - Large model, 2,000 conversations: about $120 - Total: about $160 a month

That is roughly a quarter of the cost of using the large model for everything. The split in your business may be different, so measure it.

Count the costs beyond tokens.

The model fee is rarely the whole bill. Other costs to include are:

- Search and storage. If your bot looks up help articles, you need a way to store and search them. This usually adds a small monthly cost. - Building and maintaining the bot. Developer time to set it up, connect it to your systems and fix problems. - Monitoring and review. Someone needs to check answers, especially early on. - Retries and errors. Failed requests and repeated attempts still use tokens.

These can add up to more than the model itself, especially at low volumes.

Compare cost per good answer.

A cheap model that gets answers wrong is not really cheap. Wrong answers lead to unhappy customers and extra work for your team.

The most reliable way to compare is to run your own test. Collect a set of real questions your chatbot will face. Include easy ones, tricky ones and a few that the bot should refuse or hand to a human. Run each question through the models you are considering, using the same instructions.

For a support bot, 50 to 100 questions is usually enough to see clear differences. For example, include a simple one ("How long does delivery take?"), a tricky one ("Can I return a sale item I bought abroad?") and one the bot should not answer ("Can you give me a refund right now?").

Then score the answers. Have someone who knows the subject do the scoring. A simple scale works fine, such as 0 for wrong, 1 for partly right and 2 for fully right. Record the tokens each model used too. Now you can compare cost against quality, not just price.

It also helps to compare against your current cost. If a person handles each enquiry today, estimate what that costs. Then you can see what the chatbot saves, even if it only handles some conversations on its own.

Match the model size to the task.

In our experience, the leading models from OpenAI, Anthropic and Google are all capable in business settings. The bigger cost question is which size of model each task needs. These are general starting points, not fixed rules or recommendations, so confirm them with your own testing.

- Customer support from your own content. Mid-range models are a sensible place to start testing. Check closely how well each one sticks to your source material and admits when it doesn't know. That matters most where wrong answers are costly. - Long documents, such as contracts, reports or policy manuals. These use a lot of input tokens, so input price matters most here. Watch for long-context surcharges too. Some providers charge a higher rate once a single request passes a certain size. At the time of writing, Gemini Pro's input price doubles above 200,000 tokens, and OpenAI's GPT-5.4 adds a surcharge above 272,000 tokens. Gemini and Claude models can both take in large amounts of text at once, but check where the higher rates start. - Pulling structured data, such as extracting invoice totals into a spreadsheet format. Small and mid-range models are worth testing first, which keeps costs down. OpenAI, Anthropic and Google all offer features for returning data in a fixed format. - Drafting emails and written content. Output tokens drive the cost here. Let the people who will use the drafts judge which model writes well enough, and whether a cheaper one will do. - High-volume, simple tasks, such as tagging or routing messages. The small models are usually enough: GPT mini models, Claude Haiku or Gemini Flash. At high volume, the price difference adds up fast.

Price out self-hosting properly.

Open models like Llama or Mistral have no per-token fee. That can look like a big saving, but the costs move elsewhere. You pay for powerful servers, either your own or rented from a cloud provider. You also pay for the staff time to set them up, keep them running and update them.

Self-hosting tends to make sense at very high volumes or where data must stay in-house. For most businesses starting out, a hosted service is cheaper overall. Run the numbers both ways before deciding.

Check your existing contracts.

Your existing cloud setup can affect cost. OpenAI models are available through Microsoft Azure. Claude is available through Amazon Web Services, Google Cloud and Microsoft Azure. Gemini fits naturally with Google Cloud and Google Workspace. If you already have spending commitments or credits with one of these, using it may lower your real cost. It can also make billing and security reviews simpler.

Keep costs under control after launch.

Set spending limits and alerts in each provider's dashboard. Review usage every month at first. Watch for conversations that run unusually long, as they often point to a problem worth fixing.

Avoid locking yourself in.

The models improve quickly, and prices shift. The cheapest good option today may not be the cheapest in a year. Build your chatbot so you can swap the underlying model without rebuilding everything. Keep your test questions and cost spreadsheet handy so you can rerun them whenever prices change or a new option appears.

The bottom line.

Estimating the cost of an AI assistant is less about the price list and more about how your chatbot uses it. Define the job, count your tokens, plan for a range, include the hidden costs and compare cost against quality. That approach will serve you well, whichever model you pick.