ChatGPT API cost per month for a small business is usually under $5, until one feature blows it up

There is no monthly price, and ChatGPT Plus is not the API

The ChatGPT API does not have a monthly fee. You load prepaid credit onto an OpenAI developer account and every request draws it down a fraction of a cent at a time. What you pay in a month is whatever you used.

The most common mix-up is worth clearing up first. ChatGPT Plus, the $20-a-month subscription, does not include any API access. They are two separate products on two separate bills. Plus gets you the chat app in your browser. The API is what software uses when it talks to OpenAI's models on your behalf, like a tool that drafts email replies inside your inbox or summarizes form submissions. You can have one without the other. Paying for Plus does not give you a single API credit.

Strictly speaking there is no product called the "ChatGPT API." It is the OpenAI API, and ChatGPT is OpenAI's chat app built on the same models. People search for it under the ChatGPT name, so that is the name used here.

The billing unit is the token, roughly three-quarters of a word. You pay one rate for the text you send in (input) and a higher rate for the text the model writes back (output). If tokens are new to you, we covered how per-request token pricing works separately. This page assumes you know that much and want an actual monthly number.

So the honest answer to "how much per month" is a multiplication. It depends on how many tasks you run, how long each one is, and which model you point them at. The rest of this page works through that multiplication for a real-shaped business, then shows the two things that make the number jump.

The prices you are multiplying by

OpenAI publishes prices per million tokens. These are the list prices OpenAI announced for its GPT-5 family when it launched in August 2025:

Check the OpenAI pricing page before you trust any number on this page, including these. OpenAI changes prices and renames models often, and the direction has mostly been down. The method below stays the same either way. Swap in today's rates and redo the arithmetic.

Three things in that table matter more than the headline figures:

• Output costs eight times input. A task that reads a long email and writes a two-line answer is cheap. A task that reads two lines and writes a page is not. When you estimate, count what comes out more carefully than what goes in.
• Cached input is 90% off. If many of your requests start with the same long block of instructions, OpenAI can reuse that part and bill it at the cached rate. This happens automatically for prompts over about 1,024 tokens. You do not switch it on.
• Batch jobs are half price. If something does not need an answer right now, such as tagging last month's leads overnight, OpenAI's Batch API runs it within 24 hours at 50% off. Most small-business tools do not use it, but it is worth knowing it exists.

The gap between the models is large. GPT-5 costs five times GPT-5 mini, and GPT-5 mini costs five times nano. For most office writing, mini is plenty. Nano is fine for sorting and tagging, and the full model is worth it for the occasional task that needs real judgment.

A worked month for a six-person body shop

Take a hypothetical collision repair shop, call it Dale's Collision, with six people and a steady flow of insurance work. Here is what they hand to the API in a month, with token counts that are deliberately on the generous side:

Add it up and you get about 889,000 input tokens and 201,600 output tokens across 690 tasks. That is a busy front office. Many small businesses will run less than half of this.

Now multiply by the prices above:

• GPT-5 nano: 0.889M × $0.05 + 0.2016M × $0.40 = about $0.13 a month
• GPT-5 mini: 0.889M × $0.25 + 0.2016M × $2.00 = about $0.63 a month
• GPT-5: 0.889M × $1.25 + 0.2016M × $10.00 = about $3.13 a month

That is not a typo. The whole month of back-office writing, on the mid-tier model, costs less than a cup of coffee. Even on the flagship model it comes in at a few dollars. That fits what we found when we priced the average AI task at a tenth of a cent to 3 cents.

If the numbers were this tidy, this page could stop here. But those figures leave out two things. Either one can multiply your bill several times over, and most first-page answers on Google do not mention either.

The line item you cannot see: reasoning tokens

GPT-5 and its smaller siblings are "reasoning" models. Before they write the answer you see, they can think through the problem in text you never see. That hidden thinking is billed as output tokens, at the expensive output rate.

OpenAI's own reasoning guide says reasoning tokens are not shown in the response but do count toward usage and billing. How many get used depends on a setting called reasoning effort, which runs from minimal up to high. A simple reply might use a few hundred reasoning tokens. A tricky one at high effort can use thousands, often more than the visible answer.

Here is what that does to Dale's month. Suppose each of the 690 tasks burns an extra 1,500 reasoning tokens on average. That is an illustrative figure, not a measured one, because yours depends on the task and the setting. That adds about 1.035 million output tokens:

• GPT-5 mini: $0.63 becomes about $2.70
• GPT-5: $3.13 becomes about $13.48

The work is identical. Only the hidden thinking changed, and it quadrupled the bill.

What to do about it:

• Use low or minimal effort for routine writing. Sorting an email, summarizing a form, or drafting a polite reply does not need deep thought. Save higher effort for estimates with tricky line items, or for anything where a wrong answer costs you money.
• Look at your usage dashboard after a week. If output tokens are several times larger than the replies you are actually getting, reasoning is the reason.
• If you use a tool built on the API, ask what effort level it sends. Many tools never expose this, and a tool that defaults to high effort on everything is spending your money for nothing.

Even at the reasoning-heavy end, $13 a month is not frightening for a six-person shop. The point is that you should know where the money goes, because the same blind spot gets more expensive in the next case.

One chatbot can outspend your whole back office

The expensive pattern is not "more tasks." It is long conversations. The API has no memory. Every time a customer sends another message to a chatbot, the software resends the entire conversation so far, and you pay for all of it again as input.

Say Dale puts a chatbot on the website to answer "how long will my car take" and "do you work with my insurer." Each message adds about 400 tokens of context, and some conversations run 20 turns. By turn 20 the model is re-reading 8,000 tokens just to write one more line. Across all 20 turns that adds up to roughly 84,000 input tokens for a single conversation, plus about 4,000 tokens of replies.

At 300 conversations a month, that is about 25 million input tokens and 1.2 million output tokens:

• On GPT-5 mini: about $6.30 input + $2.40 output = $8.70 a month
• On GPT-5: about $31.50 input + $12.00 output = $43.50 a month, before any reasoning tokens

So one chat widget on the flagship model costs more than ten times Dale's entire back office on the same model. Most conversations will be shorter than 20 turns, so treat this as the ceiling rather than the average. The shape still holds: cost grows with the square of conversation length, not in a straight line.

Caching helps. Most of each resent conversation is identical to the last request, so much of it can bill at the 90% cached rate. OpenAI's prompt caching documentation explains when that applies. But caching is not guaranteed on every request, and you should not budget as if it is.

The same trap shows up anywhere context piles up. Pasting a 40-page PDF into every request, or letting an "agent" loop through a dozen steps that each re-read everything before them, both do it. If your bill ever surprises you, look for a feature like this first. It is almost never the plain one-shot tasks. Those are the cheap, boring workhorses we described in most small-business AI automation being plain rules with one or two smart steps.

How to find your own number and put a ceiling on it

You do not need to estimate tokens by hand. Run a week of real work and let OpenAI count for you.

1. Create a separate project for the business in the OpenAI platform and make an API key for it. That keeps business usage apart from anything else you experiment with.
2. Load a small amount of credit and turn auto-recharge off. Because the API is prepaid, a balance of $10 with auto-recharge off means the most you can ever lose to a runaway feature is $10. When it runs out, requests stop.
3. Set a monthly budget alert on the project so you get an email well before you hit your limit.
4. Run one normal week, then open the Usage page. Multiply the week's spend by about 4.3 for a monthly figure.
5. Look at the split by model and by input versus output. If one model or one day dominates, that is your chatbot or your reasoning setting. Fix that one thing rather than cutting AI everywhere.

Once you know your real monthly number, use it as a yardstick for every tool that sells you "AI." Plenty of software charges $20 to $100 a month for an AI add-on or bundles of AI credits. Convert their price into tasks. If your real API usage is $3 a month and the add-on is $50, the rest is markup. That can be a fair price for convenience, but it is a choice you should make knowingly.

This is the reason Fynl's AI runs on your own OpenAI or Claude key. You pay OpenAI their list rate directly, the $49-a-month portal fee does not change with usage, and the key stays yours if you ever leave. It does not handle SMS or phone calls yet, so if a text-message assistant is the feature you want, you will need something else for that.

If you are weighing OpenAI against Anthropic, the numbers come out in a similar range. We did the same exercise in what Claude actually costs a small business. Pick on output quality for your kind of writing, not on a few cents of difference.

Questions people also ask

Sources

• OpenAI API pricing
• OpenAI: Introducing GPT-5 for developers
• OpenAI docs: Reasoning models
• OpenAI docs: Prompt caching
• OpenAI docs: Batch API
• ChatGPT plans and pricing