For a typical small-business automation, the ChatGPT API costs from a few cents to a few hundred dollars a month in OpenAI usage fees. Sorting and drafting replies to 5,000 emails a month costs about $1.90 on gpt-6-luna, $38 on gpt-6-sol or $190 on gpt-6-astra. The automation platform that runs it often costs more than the AI.
Two things catch people out. A ChatGPT Plus, Pro or Business subscription doesn't include any API use: the API (the connection that lets other software send work to OpenAI's models) is billed separately, per token, on OpenAI's developer platform, even if you log in with the same email address. And the model fee is rarely the biggest line. At low volumes the platform subscription dominates; at high volumes, the time spent checking what the AI wrote does.
What a single API call is charged for
OpenAI prices its models per million tokens. A token is a chunk of text, usually part of a word; a workable rule of thumb is that 1 million tokens is roughly 750,000 English words, so 1,000 words is about 1,330 tokens. Every call your automation makes is charged on up to four meters:
- Input tokens: everything you send. That means your instructions, any examples, and the email, form or document being processed.
- Cached input tokens: the repeated opening part of your prompt. On GPT-5.6 and later models, OpenAI caches an identical prefix of at least 1,024 tokens automatically and charges a tenth of the normal input price when it is reused within 30 minutes.
- Output tokens: what the model writes back. This includes hidden reasoning tokens: OpenAI's reasoning guide says they are billed as output even though you never see them.
- Tool calls: if the model searches the web, OpenAI charges $10 per 1,000 calls plus the search content at normal token rates; its file search tool is $2.50 per 1,000 calls.
Most business automations only use the first three. Output is the expensive meter, at five to eight times the input price on current models, so a long answer costs far more than a long question.
OpenAI's token prices, per million, in September 2026
These are the list prices on OpenAI's API pricing page for standard processing. The GPT-6 series launched in September 2026; the older series stay available at their own prices.
| Model | Input | Cached input | Output | Typical automation job |
|---|---|---|---|---|
| gpt-5-nano | $0.05 | $0.005 | $0.40 | Simple labelling, yes/no checks |
| gpt-6-luna | $0.10 | $0.01 | $0.50 | Sorting, extracting fields, short drafts |
| gpt-5.6-luna | $0.20 | $0.02 | $1.20 | The same jobs on the previous series |
| gpt-6-sol | $2 | $0.20 | $10 | Customer-facing drafts, summaries of longer documents |
| gpt-5.6-terra | $2 | $0.20 | $12 | Mid-weight drafting on the previous series |
| gpt-5.6-sol | $4 | $0.40 | $20 | Promotional price, at least to 21 Nov 2026 |
| gpt-5.5 | $5 | $0.50 | $30 | Older flagship |
| gpt-6-astra | $10 | $1 | $50 | Hard reasoning over long technical material |
For sorting, extracting and drafting, the cheap tiers are usually enough. Start with gpt-6-luna, and move a job up only when it fails a test on your own examples.
Measuring one run of your automation
Price one run properly before multiplying anything. Take 20 real examples of the input (emails, forms, delivery notes), write the instructions, and count. Here is the prompt an illustrative import-export business uses to triage its shared inbox:
You sort incoming emails for a small import-export business.
Return JSON only, with these fields:
category: one of quote_request, order, shipment_query,
invoice, complaint, other
urgent: true or false
customer_ref: the customer's PO or reference, or null
container_or_awb: container or air waybill number, or null
draft_reply: under 120 words; empty if category is invoice
Urgent means: goods are held at a port or airport, the sender
states a deadline within 48 hours, or the sender is escalating.
Never promise dates, times, prices or customs outcomes.
[Three worked examples follow, about 700 words]
EMAIL:
{{email_body}}An illustrative output for a customer chasing a delayed container:
{"category": "shipment_query", "urgent": true,
"customer_ref": "PO-44817", "container_or_awb": "MSKU7784512",
"draft_reply": "Hi [first name], thanks for flagging this. I'm
checking the status of container MSKU7784512 with our forwarder
now and will update you by 3pm today."}The sorting and extraction are right. The draft is not: "by 3pm today" is exactly the kind of time promise the instructions forbade, and on a busy day nobody may be free by 3pm. Two fixes: add a check step that holds any draft containing a time or date for a person to approve, and put one example in the prompt of a correct reply to a delay ("I'll update you as soon as I hear back"). Rules work better with an example beside them.
Now the count. The fixed part (instructions plus examples) is about 1,100 words, or roughly 1,500 tokens. A typical email adds 220 words, about 300 tokens, so each call sends around 1,800 input tokens. The JSON and draft come back at about 250 tokens, plus reasoning at a low effort setting; allow 400 output tokens in total. One run then costs:
- gpt-6-luna: 1,800 x $0.10 per million + 400 x $0.50 per million = about $0.00038
- gpt-6-sol: 1,800 x $2 per million + 400 x $10 per million = about $0.0076
- gpt-6-astra: 1,800 x $10 per million + 400 x $50 per million = about $0.038
Check your estimate against reality after the first day. Every API response reports the tokens used, including a separate count of reasoning tokens, and the usage page on OpenAI's platform totals them.
Email triage at an import-export firm: 500, 5,000 and 50,000 a month
Multiply the per-run cost by volume and the three models separate sharply. These are monthly model fees for the same triage automation, at standard prices with no caching:
| Emails a month | gpt-6-luna | gpt-6-sol | gpt-6-astra |
|---|---|---|---|
| 500 (a two-person office) | $0.19 | $3.80 | $19 |
| 5,000 (a busy shared inbox) | $1.90 | $38 | $190 |
| 50,000 (several inboxes plus web forms) | $19 | $380 | $1,900 |
Caching changes the top row most. At 50,000 emails a month, calls arrive every minute or so during the working day, so the 1,500-token fixed prefix stays in the 30-minute cache and is billed at a tenth of the input price. On gpt-6-sol that cuts the input cost per run from $0.0036 to about $0.0009, and the monthly bill from $380 to about $245. At 500 emails a month, gaps between emails often run past 30 minutes, so treat caching as a bonus rather than something to budget on.
The practical answer for most small firms: gpt-6-luna for sorting and extraction, gpt-6-sol for any reply a customer will read, and gpt-6-astra only where a cheaper model has visibly failed on your test set.
Why the automation platform usually costs more than the AI
The model does the thinking, but something has to watch the inbox, pass each email to OpenAI, save the draft and log the result. On Zapier, a trigger is free and each successful action step counts as one task. The triage flow above has three actions (the OpenAI step, create a draft, add a spreadsheet row), so every email uses three tasks. Here are Zapier's published Professional tiers for those volumes alongside the model fee:
| Emails a month | Tasks used | Zapier Professional tier | Zapier price (annual / monthly billing) | gpt-6-luna fee |
|---|---|---|---|---|
| 500 | 1,500 | 2,000 tasks | $49 / $73.50 | $0.19 |
| 5,000 | 15,000 | 20,000 tasks | $189 / $283.50 | $1.90 |
| 50,000 | 150,000 | 200,000 tasks | $769 / $1,149 | $19 |
At every volume the platform costs more than the tokens on a cheap model, by a factor of 40 or more. Make is cheaper per step: its entry plan is about $9 a month for 5,000 credits, with one credit per module action, but a scheduled trigger also uses a credit each time it checks for new mail, even when nothing has arrived. Checking every 15 minutes uses about 2,880 credits a month on its own. If your volumes are heading towards the top row, working out when Zapier gets too expensive is worth an hour before you build.
One more platform detail. Zapier's own AI by Zapier step uses 1, 3 or 5 tasks per run depending on the model tier you choose, or 1 task if you connect your own API key. Paying OpenAI directly through your own key is often cheaper than paying for the same work in extra tasks. Adding AI steps to Zapier walks through both set-ups.
Then there's the third cost: people. If someone approves each draft at 30 seconds a time, 500 emails is about four hours a month, 5,000 is about 42 hours and 50,000 is over 400. At the top volume you cannot review everything, so the design question becomes which categories may send without approval. Invoices filed to a folder, perhaps; complaints, never. Human approval steps in AI automations covers how to split them.
Where API bills balloon: four realistic mistakes
The automation that answered itself
An illustrative care agency set up an automation that acknowledged every email to its enquiries inbox. One evening it acknowledged a message from a supplier whose mailbox also replied automatically, and the two systems answered each other all night. With Zapier checking for new mail every two minutes, that was about 300 runs by morning, and nearly half the month's tasks gone. It showed up as a task-limit warning, not as a bad email, because every reply looked perfectly polite. The fix is a filter as the first step (skip automatic replies, no-reply addresses and anything from your own domain), a cap on runs per sender per day, and a hard spend limit, covered below.
Sending the whole PDF to find three numbers
An illustrative testing laboratory used the API to pull the sample ID, test code and result from each report, and sent the whole 30-page PDF text every time: around 15,000 words, or 20,000 input tokens, per report instead of about 1,500. That's more than ten times the cost per run, and the extraction got worse because the answer was buried. Send only the page or section that holds the fields, and test whether the cheaper model copes once the noise is gone. It usually copes better.
The flagship model on a sorting job
An illustrative packaging supplier built its quote-request sorter on gpt-6-astra because it was "the best". At 3,000 emails a month with prompts similar to the one above, the model fee was about $114. A test on 50 of its own past emails showed gpt-6-luna gave the same category on 49 of them, and the one it missed was ambiguous to people too. Switching cut the fee to under $1.20 a month. Test with your own examples; the job decides the model, not the leaderboard.
Reasoning left on maximum
An illustrative spare-parts manufacturer copied a template that set reasoning effort to its highest setting for a simple part-number lookup. Each call produced a short visible answer but thousands of hidden reasoning tokens, all billed as output. The response objects showed it plainly in the reasoning token count; nobody looked until the invoice arrived. OpenAI's reasoning settings run from none and minimal up to xhigh and max (not every model supports every value). Use the lowest setting that passes your test set, and set a maximum output length on each call as a ceiling.
Paying less: cached instructions, overnight batches, smaller models
Order your prompt for the cache. Put the fixed instructions and examples first and the changing email last, and keep the fixed part identical on every call. Even a date stamp at the top breaks the match. The prefix must reach 1,024 tokens to be cached at all, so for a short prompt caching won't apply, and that is fine: short prompts are cheap anyway.
Batch anything that can wait a day. OpenAI's Batch API gives a 50% discount in exchange for results within 24 hours, often sooner. Picture the spare-parts manufacturer rewriting 4,000 catalogue descriptions on gpt-6-sol, with about 600 input tokens and 250 output tokens each:
Input: 4,000 x 600 = 2.4 million tokens x $2 = $4.80
Output: 4,000 x 250 = 1.0 million tokens x $10 = $10.00
Standard price: $14.80 Batch price (50% off): $7.40Either way it's cheaper than one hour of the person who will check the results, which is the real cost of that job.
Route by difficulty. Let the cheap model sort everything, and pass only the hard cases up. If the import-export firm's 5,000 emails a month include 500 complaints, drafting just those on gpt-6-sol and everything else on gpt-6-luna costs about $1.90 plus $3.80, around $5.70 a month instead of $38 for the whole inbox on gpt-6-sol.
Ask for less output. A 120-word limit on drafts and JSON-only answers do more for the bill than any discount, because output is the expensive meter. OpenAI's structured outputs feature, which makes the model follow a fixed format, also makes results more consistent; turning the temperature setting down is no longer a reliable way to do that on current models.
Capping spend before you switch the automation on
A cheap automation can still produce an expensive month if it loops or a prompt changes. OpenAI's spend limits guide describes the controls, and they take ten minutes to set:
- One project per automation. Create a separate project on the platform for each automation, with its own API key, so usage is reported separately and one runaway flow can't hide inside another.
- Set a monthly spend limit on the project. In the project's settings, open Limits and edit the spend limit. Put it at roughly three times your expected monthly cost.
- Turn on Enforce a hard limit. Alerts alone don't stop spending. With enforcement on, requests fail with a 429 error once the limit is reached.
- Decide what happens on failure. When the API refuses a call, your automation should pass the email to a person or a holding folder, not retry in a loop.
- Add an organisation-wide limit as a backstop across every project.
A worked version: the import-export firm expects about $2 a month on gpt-6-luna for its triage flow, so it sets a $6 hard limit on that project and a $50 limit across the organisation. If a runaway loop like the care agency's ever hit this project, the OpenAI spend would stop at $6 and the emails would queue for a person.
When a ChatGPT seat is the better buy than the API
The API is for work that runs without a person starting it. When someone sits down and works with the AI, a seat is simpler and usually better value. Before deciding the API is the answer, do the sum in working out an AI automation's return before you build it, because the build and testing time often outweighs a year of token fees.
| Job | Better on | Why |
|---|---|---|
| Owner drafting proposals and reports | ChatGPT Plus at $20 a month | Interactive back-and-forth, no build needed |
| A team sharing projects and files | ChatGPT Business at $25 a seat monthly or $20 annually, minimum 2 seats | Admin controls, no training on business data by default |
| Every incoming email sorted and drafted | API through Zapier, Make or similar | Runs on a trigger, unattended, per-token billing |
| 4,000 product descriptions once a year | API via the Batch API | Half price, results next day |
| Weekly look at the sales spreadsheet | A seat | A person needs to ask follow-up questions |
Most small businesses end up with both: a handful of seats for the people, and one or two API automations costing less per month than a single seat. If your API estimate comes out above a few hundred dollars a month, recheck the model choice and the prompt length before you recheck the budget.
Questions about paying for the OpenAI API
Is the ChatGPT API the same as a ChatGPT subscription?
No. ChatGPT Plus, Pro and Business are subscriptions for people using the chat app, and none of them includes API use. The API is billed separately, per token, on OpenAI's developer platform, with its own billing settings, keys and limits. You can use both, and many businesses do: seats for staff, the API for automations that run unattended.
Does OpenAI train its models on data sent through the API?
Not by default. OpenAI doesn't use API data for training unless you opt in. That doesn't remove your own duties: check what personal data your automation sends, keep it to the fields the job needs, and record the processing in whatever data-protection paperwork your business keeps.
Do I need a developer to use the OpenAI API?
Not for most small automations. Zapier, Make and similar platforms have OpenAI steps that only need an API key pasted into a connection. You still need someone comfortable writing clear instructions, testing on real examples and reading the usage page, which is closer to an operations job than a programming one.
How do I see what each automation costs?
Create a separate project on OpenAI's platform for each automation, with its own API key. Usage and spend are then reported per project, and you can set a spend limit on each one. Check the figures weekly for the first month, then monthly once the numbers settle.
Further reads
- How to Build Your First AI Automation in Make, Step by Step — Build a first AI automation in Make, with the API key step included.
- Zapier vs Make vs n8n for AI Automation: Which Fits Your Business? — Choose the platform, which is often the bigger monthly cost.
- How Much Does It Cost to Automate One Workflow With AI? — The whole cost of one workflow, including build time.
- AI Maintenance Costs: What You Pay After an Automation Goes Live — What keeps costing money once the automation is live.
- Per-Seat vs Usage-Based AI Pricing: Which Costs Less for You? — When usage pricing beats paying per seat, and when it doesn't.
- How to Stop Zapier and Make Automations Breaking Silently — Alerts that catch a failing or looping automation early.
- How Much Does AI Cost a Small Business in 2026? — Itemised 2026 AI budgets for a sole trader, a six-person office and a 20-person firm, with the charges that grow after month three.
- What Is an API? Why It Matters When You Buy Software — A plain-English explanation of APIs for people buying software, with the vendor questions and pricing traps that decide whether tools can connect.
- How Much Does a Custom AI Assistant Cost a Small Business? — Three ways to get an AI assistant that knows your business, what each costs up front and every month, and the upkeep that most quotes leave out.
- RAG vs Custom GPT vs Fine-Tuning: What Does Your Business Need? — Three levers compared: RAG changes what the AI knows, fine-tuning changes how it behaves, and custom GPTs are retiring. A costed dental practice decision.
- How to Set Spending Limits and Alerts on Pay-As-You-Go AI Tools — Where the spend limits live in OpenAI, Anthropic, ChatGPT, Claude, Zapier, Make and n8n, and how to choose caps that stop a runaway without breaking bookings.
- Zero Data Retention: What It Means When You Choose an AI Tool — What a zero data retention promise really deletes, the exceptions OpenAI, Anthropic and Google keep, and the vendor questions that expose weak claims.
- Open-Source vs Paid AI Models: What Small Businesses Should Know — How to choose between open-weight and paid AI models: what the licences allow, how prices compare, and a tested way to decide for your own documents.
- How to Read AI Software Pricing: Seats, Credits, and Usage Fees — The six pricing units AI software uses, the fine print that changes the total, and a sports shop turning three pricing pages into one monthly figure.
- What Does It Cost to Integrate AI Into Your Existing Software? — Four ways to add AI to software you already run, what each costs to build and run, and a letting agency's three quotes compared per request.
- AI Translation vs Human Translators: What Documents Really Cost — What a 2,000-word document costs by AI, by AI plus a bilingual reviewer, and by a human translator, with a tour operator's year costed three ways.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI API pricing page (developers.openai.com), OpenAI guides on prompt caching, reasoning, the Batch API and spend limits; Zapier pricing page (task tiers); Make pricing page. All checked September 2026.