ChatGPT API cost: how to estimate your monthly bill

Subscription vs API

ChatGPT Plus, Business and Enterprise are per-person subscriptions for people using ChatGPT in the app. The OpenAI API is pay-as-you-go: your software sends text to a model and you pay for the tokens used.

Your API bill has nothing to do with how many ChatGPT seats you have, so budget for them separately.

The quick estimate

Work out four numbers for each automation: how many requests a month, how many tokens go in per request, how many come out, and which model it uses. Multiply input and output tokens by the model's rates per million tokens and add them together. [Confirm current OpenAI model rates before publishing.]

Add a margin of 20 to 30 per cent for retries, testing and growth. If the number surprises you, try the same sum with a smaller model. For sorting, tagging and short replies, the smaller models are often good enough at a fraction of the price.

What pushes the cost up

Long chat histories sent with every message, whole documents where a section would do, asking for long answers when a short one works, and retry loops that fire on every error.

Reasoning models also produce extra internal tokens while they think, which are billed as output. Use them where the reasoning genuinely helps.

How to keep it down

Set monthly budget limits and alerts in your OpenAI account, use separate projects or keys for each automation, and check usage weekly while a new automation beds in. Then apply the usual token optimisation steps: trim prompts, cache repeated context, use the batch service for work that can wait, and choose the smallest model that gets the job right.

How OpenAI API pricing works

People often search for the ChatGPT API, but there isn't a separate ChatGPT API. What you're paying for is the OpenAI API, which gives your software access to the same family of OpenAI models that power ChatGPT. You sign up on OpenAI's developer platform, create an OpenAI API key, add a payment method or buy API credits, and pay for what you use.

OpenAI API pricing is per token. A token is a small chunk of text, roughly three quarters of an English word on average. Each OpenAI model has an input price and an output price, usually quoted per 1M tokens. Output rates are higher than the input rate on most models. [Insert current OpenAI model rates per 1M tokens.]

So the core sum is the same for every model: input tokens multiplied by the input rate, plus output tokens multiplied by the output rate. The model price you choose matters more than almost anything else, because the gap between the cheapest and most expensive OpenAI model is large.

Input tokens

Input tokens are everything you send: system prompts, tool definitions, chat history, documents, any image input and the user's message. For most business automations, the repeated system prompts and context make up the bulk of the input.

Output tokens

Output tokens are everything the model sends back. With reasoning models, this also includes the reasoning tokens the model uses while it thinks. You don't see those in the response, but they're billed as output, so they can make a short answer surprisingly expensive.

OpenAI models and what they cost

OpenAI releases new models often, and earlier models usually stay available for a while at their own rates. Rather than quote a price sheet that will be out of date by next month, here's how the range breaks down. Check OpenAI's pricing page for current names and API prices. [Confirm current OpenAI model names and rates before publishing.]

• Flagship models: the most capable general models, with the highest price per million tokens.
• Smaller and mini models: much cheaper, and often good enough for sorting, tagging, extraction and short replies.
• Reasoning models: built for complex, multi-step problems. They use extra reasoning tokens, so cost more per task.
• Specialist models: for images, audio, transcription, embeddings and real-time voice, each priced in its own way.
• Fine tuned models: versions of base models trained on your own examples, with separate training and usage rates.

Model choice is your biggest lever

Using the same model for every job is the most common reason an OpenAI bill is higher than it needs to be. Test the smallest model that might work first, and only move up when the results aren't good enough. Many businesses end up with two or three models, each matched to a different automation.

Reasoning effort settings

Many reasoning models let you set the reasoning effort. Lower effort uses fewer reasoning tokens and costs less. Higher effort can help with hard analysis but costs more per request. Start low and raise it only where you can see the quality improve.

Ways to pay less: cached input, Batch API and processing tiers

OpenAI offers several ways to lower the rate on the right kind of work. Each one has rules, so check how they apply to the model you use.

Cached input

When the start of a prompt repeats across requests, such as a long system prompt or a set of tool definitions, OpenAI can reuse it and charge cached input at a lower rate than the standard input price. On many models this happens automatically, as long as the repeated part sits at the start of the prompt and stays exactly the same. [Insert current cached input rates.]

The Batch API

The Batch API lets you send a large set of requests and collect the results later, at a lower rate than standard requests. It suits work that can wait, such as bulk document processing, data clean-up, tagging or overnight reports. It isn't right for live chat. [Insert current Batch API discount and turnaround window.]

Priority processing and other tiers

For some models, OpenAI offers priority processing for faster, more consistent responses at a higher price, and lower-cost options for work that doesn't need speed. Regional processing, where data is handled in a particular region, can also carry a different rate. [Confirm current processing tiers and regional processing pricing.]

Long context pricing

Some models charge a different rate once a request goes beyond a certain length. If you regularly send very long documents, check whether long context pricing applies, and trim what you send to the parts that matter. [Confirm current long context pricing rules.]

Built-in tools and extra charges

The Responses API includes built-in tools such as web search, file search and code interpreter. These can be very useful, but several carry charges on top of normal tokens, for example per tool call, per amount of storage or per session. Tool definitions and tool results also add to your input tokens. [Insert current built-in tools pricing.]

Function calling, where the model calls your own systems, doesn't have a separate fee, but every function definition you send counts as input tokens. Agents that loop through many tool calls can use far more tokens than a single question, so set a limit on how many steps they can take. Deep research style features can be especially token heavy.

Usage tiers, limits and your OpenAI bill

New OpenAI accounts start on a default tier with lower rate limits. As you spend more and your account ages, you move up usage tiers, which raises how many requests and tokens you can use per minute. Usage tiers affect limits, not the price per token. [Confirm current usage tier rules.]

In your OpenAI account, you can set a monthly usage limit and email alerts at set thresholds. Use projects to separate each automation, give each project its own OpenAI API key, and you'll be able to see exactly where your OpenAI API spend is going.

Azure OpenAI pricing

If your business already uses Microsoft Azure, you can buy access to OpenAI models through Azure OpenAI Service instead. Azure OpenAI pricing is set by Microsoft, billed on your Azure invoice and offers options such as provisioned throughput units for predictable, high-volume work. It's worth comparing both if you're already on Azure. [Confirm current Azure OpenAI pricing.]

ChatGPT subscription or OpenAI API: which do you need?

If your team wants to chat with AI, draft documents and analyse files, ChatGPT's business plans or enterprise plans are usually the right fit, priced per person with a predictable monthly bill. If you want AI running inside your own software, website or automations, you need the OpenAI API.

Many small businesses need both. Budget for them separately, because a ChatGPT subscription doesn't include API credits, and API usage doesn't count against anyone's ChatGPT plan. The best results usually come from combining all three: the AI model, people who know the work, and your own business data feeding into it.