Average AI API cost runs from a tenth of a cent to 3 cents per task

What one ordinary task actually costs on four common models

An AI API bills you per token, about three-quarters of a word. Output costs more than input, usually four to eight times more per token. If that is new to you, how per-request token billing works covers the basics. This page is about the number you came for.

A single "average cost per call" means very little. A short question and a 10-page contract are both one call, and they differ in price by 10x or more. The useful number is cost per task. Below are three tasks a small business does every week, with token counts we estimated from typical lengths. A thread of a few emails is about 1,500 tokens in, and a reply is about 400 out.

The rates are the published per-million-token list prices for each model, input / output: Claude Sonnet $3 / $15, Claude Haiku 4.5 $1 / $5, GPT-5 $1.25 / $10, GPT-5 mini $0.25 / $2. Providers change these. Check the pricing pages in the sources before you rely on them.

So the honest answer to "average API cost" is about one cent per routine task on a top mid-range model. On the small, fast models it is about a tenth of that.

One catch on GPT-5: it is a reasoning model. The hidden "thinking" tokens it produces are billed as output, so real costs can run two or three times the figure above unless you set reasoning effort to low. Claude works the same way if you turn on extended thinking.

A real month: a kitchen remodeler's bill

Take Dana, who runs a three-person kitchen remodeling business. She uses AI for three jobs. She drafts about 20 email replies a working day, which comes to 440 a month. She summarizes about 30 supplier spec sheets or contracts a month. She writes the description section of about 40 quotes.

Multiply the table by those counts and here is her monthly API bill on each model:

• Claude Sonnet: $4.62 emails + $0.95 PDFs + $0.46 quotes = about $6.02
• GPT-5 (before reasoning tokens): about $3.32, realistically $5 to $9 with default reasoning
• Claude Haiku 4.5: about $2.01
• GPT-5 mini: about $0.66

For comparison, one ChatGPT Plus seat is $20 a month, and Dana has three people. Three seats cost $60 a month, against single-digit dollars for API usage at this volume. The per-seat ChatGPT pricing and Claude's full cost range for a small business go deeper on each provider. The point here is that the gap between models matters more than the gap between providers.

What is not true: that the cheap models are useless. For email drafts and quote wording, Haiku and GPT-5 mini are usually fine. Save the bigger model for contracts and anything where a wrong detail costs you money.

Why your number can be ten times higher than these

The table assumes each task starts fresh. Most overspending comes from four habits.

Long chat threads. Every new message in a conversation re-sends the whole history as input. Say each turn adds 500 tokens. By turn 20 you are sending 10,000 tokens just to ask one more question. Over those 20 turns you send about 105,000 input tokens in total, which is about $0.32 on Sonnet. Twenty separate fresh prompts would send 10,000 tokens, about $0.03. That is 10x the cost for the same 20 questions. Start a new chat when the topic changes.

Attaching whole documents when you need one page. A 50-page PDF is roughly 40,000 tokens before you have asked anything.

Images and scans. Photos of receipts or job sites are converted to tokens too, often more than the text they contain.

Automations that run on everything. Suppose a rule summarizes every incoming email, newsletters and spam included. It can quietly run thousands of times a month. Filter what triggers it.

None of these change the per-token price. They change how many tokens you send.

How to find your own average in one week

Our numbers are estimates built from typical lengths. Yours will differ, and you can get the real figure quickly.

1. Create an API key in the Anthropic Console or the OpenAI platform. Set a monthly spending limit of $10 or $20 so a mistake cannot run away.
2. Use AI normally for one week and keep a rough tally of tasks: emails, summaries, quotes.
3. Open the usage page in the console. Both providers show spend by day and by model.
4. Divide the week's spend by your task count. That is your real cost per task. Multiply by four for a monthly estimate.

If the number surprises you, check the four habits above before switching providers.

When a tool runs on your own key, that console figure is the whole AI bill, with no markup layered on top. Fynl works this way. You plug your Claude or ChatGPT key into the CRM, email, and quote tools, and the usage shows up in your own Anthropic or OpenAI dashboard. Many all-in-one tools instead bundle AI "credits" into the subscription. It is worth checking what those credits work out to per task against the table above.

Questions people also ask

Sources

• Anthropic API pricing
• OpenAI API pricing
• OpenAI reasoning models guide (reasoning tokens billed as output)