ChatGPT free API: what is actually free, what is billed, and the three honest routes
The app and the API are separate products with separate billing. Here is what each one gives you free, how to run the API on a small prepaid balance, and when a local model is the better answer.
Published by AI Knows Us · Updated 13 October 2026
The ChatGPT app has a free tier you can use in a browser. The OpenAI API, which is the thing you call from code, is a separate paid product billed on usage, and a ChatGPT subscription does not include API calls. So "free ChatGPT API" resolves into three honest routes: use the free ChatGPT app by hand for one off work, put a small prepaid balance on an API account and pay per token for automation, or run an open weight model on your own machine where there is no per call bill at all. This page sets out each route, what it costs you in effort rather than money, and how to pick.
The two products people confuse
They have different front doors, different billing and different terms.
| ChatGPT, the app | OpenAI API | |
|---|---|---|
| How you use it | Browser, desktop and mobile apps, by hand | HTTP requests from your own code |
| Free access | A free tier exists, with limits on the newest models | Billed on usage, paid from credit on the account |
| Billing unit | Monthly subscription for the paid plans | Per token, in and out, per model |
| Does one include the other | No. API usage is billed separately | No. Credit does not cover app subscriptions |
| Where the current terms live | openai.com/chatgpt/pricing | platform.openai.com/docs/pricing |
Both pages were reachable on 10 October 2026, and both are the source of truth for their own numbers. Rates and plan contents on a product like this move several times a year, which is why this page sends you to the vendor's own page for every figure instead of printing one that goes stale.
Route one: the free app, used deliberately
If what you need is twenty answers a week, the free ChatGPT tier in a browser does the job and costs nothing. It suits drafting, summarising a document you paste in, and checking how an assistant answers a question about your market. It does not suit anything scheduled, anything that must run over a list of five hundred inputs, or anything whose output has to land in a database.
The practical ceiling is your own time. The moment you find yourself copying the same prompt in by hand for the fourth day running, the hours you are spending are larger than the API bill would have been, and it is time to move to route two.
Route two: the API, with a small prepaid balance
This is the route most people actually want when they search for a free API. Billing is per token, so a low volume job is a very small bill rather than a subscription, and you control the ceiling by how much credit you load.
The setup, in order:
- Create an account on the API platform, which is separate from your ChatGPT login even when the email is the same.
- Add credit to the organisation. Calls fail with a quota error until there is a balance, and that error is the single most common reason people conclude the API is broken when it is not.
- Create an API key and store it as an environment variable. Keys belong in the environment, never in client side code and never in a repository.
- Run the request from the quickstart, confirm you get a response, then swap in your own prompt.
- Read the rate limit guide before you loop over a list, and build in a retry with backoff.
- Set a usage limit on the account so a runaway loop stops itself.
Three choices control the bill more than anything else. Pick the smallest model that passes your own test rather than the largest available, because capability and price both scale with model size and most production jobs do not need the top model. Keep the prompt short, since you pay for input as well as output. Cache repeated work, so a result you already have is never bought twice.
The model list states what each model is built for, and the data controls guide states how API inputs and outputs are handled, which is the page to read before you send anything belonging to a customer.
Route three: an open weight model on your own hardware
Several capable models are published under open weights and run locally through a local runner such as Ollama or a comparable project. There is no per token charge at all. You pay in hardware, setup and the quality gap between a local model and a frontier one, and for classification, extraction, tagging and summarising, that gap is small enough that plenty of teams run these jobs locally and keep their frontier spend for the work that needs reasoning.
This is the right route when the data must not leave your network, when volume is high and the task is simple, or when you want a fixed cost instead of a usage one. It is the wrong route when you need the newest capability on day one.
Several other API providers also publish free tiers for their own models. Those terms sit on each provider's own pricing page, so check them there on the day you decide.
Picking in one line
- A handful of answers, by hand: the free ChatGPT tier.
- Automation at low volume: the API with a small prepaid balance and a usage cap.
- High volume, simple task, or private data: an open weight model you host.
- Newest capability, production scale: the API on a paid usage tier, with caching and a smaller model for the easy steps.
This page covers the free and low cost routes to calling ChatGPT style models programmatically. For how OpenAI's own documentation is laid out and which pages matter to a site owner, see OpenAI documentation explained.
Where people get stuck on ChatGPT free API
Is the ChatGPT API free?
The API is billed on usage, paid from credit on the account. The free access sits on the ChatGPT app rather than on the API, and the two are billed separately.
Does ChatGPT Plus include API access?
No. A subscription covers the app. API calls are billed from the credit on your API organisation, and you add that separately.
Why do I get a quota error on my first API call?
Almost always because the organisation has no credit on it yet, or because the key belongs to a different organisation from the one holding the balance. Load credit, confirm the key's organisation, then retry.
What is the cheapest way to run a few thousand calls?
Use the smallest model that passes your own accuracy check, trim the prompt, batch where the API supports it, and cache results by input so repeated inputs are answered from your own store.
Can I use the free ChatGPT app for automation?
Use the API for automation. It is the supported route, it has documented rate limits and retries, and it keeps you inside OpenAI's terms of use.
Which figures should I plan a budget from?
Take every rate from the vendor's own pricing page on the day you build the model, then rerun the sum monthly, so your budget is always based on current rates.
Free route or paid, the buyer's answer is the same test
Whichever route you choose, the model is also the thing your buyers ask. Put their questions to it with no brand name in the prompt and record which businesses it names. That gives you a visibility reading at no cost, and it is a useful first job for a free key or a local open model. The output differs from the page you published, since the model answers from reviews, roundups and forum threads as well as sites, and it differs by model. AI Knows Us takes the same reading across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews and lists the sources.