A hundred-dollar subscription is not an AI budget

In conversations with business owners I sometimes hear something like this: “but I pay $20 or $100 a month for Claude Code” (an AI coding agent), followed right away by surprise when I say that the cost of AI in their product, or in the process they want to automate, is hard to predict in advance. Many people assume it is enough to install the same tool on a server, log in to their subscription account, and the company will run for a flat fee. I understand where this comes from, because from the outside you see the same AI model in both cases, but in practice these are two different services: a subscription for a person, and access through the API (the way programs connect to a model, billed for every use). I use Claude Code as the example because I use it myself, but at OpenAI, xAI and Google it looks very similar.

A Claude Code subscription is for the person at the keyboard

I use a Claude Code subscription every day for my own work, which is exactly what it is for (a personal preference, not a recommendation: ChatGPT with Codex, Grok and Gemini have similar plans). For a flat fee (at Anthropic, $20 a month for the Pro plan, $100 for Max and $200 for Ultra, with Max and Ultra also including a monthly pool of API credits, $100 or $200, after which you pay for tokens at the standard price list) you get a usage limit that on the Pro and Max plans resets every five hours and every week, is shared between the Claude chat and Claude Code, and once it runs out you can wait or pay for extra usage at standard API rates. Anthropic also notes that how fast you reach the limit depends, among other things, on the model, that the company may impose additional limits, and that prices and plans may change at its discretion.

For comparison, in the Claude Code documentation, on the page about costs, Anthropic says that the same agent billed through the API costs, across company deployments, an average of about $13 per developer per day on which they use it, and about $150-250 a month, adding that these costs vary widely. For one person at the keyboard a subscription is convenient, but for a product meant to serve customers at any hour of the day, a limit changed at the provider's discretion seems to me a weak foundation.

What the terms allow for automation and building products

The second matter is the terms of use, and Anthropic's Consumer Terms of Service say that, except when accessing the services through an API key or where the company explicitly permits it, you may not use them “through automated or non-human means, whether through a bot, script, or otherwise”, and that you may not make your account available to other people. The Claude Code documentation on legal and compliance adds that the limits on the Pro, Max and Ultra plans assume “ordinary, individual usage”, that people building products or services should connect with an API key, that Anthropic does not let them serve their users' requests through a subscription account, and that it may enforce these rules without notice. An agent on a server, logged in to someone's subscription account and serving the company's customers, is, as I understand these clauses, exactly the case they describe. Products for your own customers come under separate commercial terms and separate API billing (that the API is billed separately from the subscription, Anthropic says plainly in its help center).

It is similar at the other providers: OpenAI bills ChatGPT and the API separately, Google says that direct use of the Gemini API is paid for outside the plans, and xAI's API is paid for in the API console, usually prepaid (you top up the account first), independently of The X Subscription. I don't want to prejudge anything here, but before you base a company on a subscription, it is worth reading the terms of your plan, because they decide what is allowed, whatever happens to work technically.

How API cost is counted: input and output tokens

With the API you don't pay for a month but for tokens, the pieces of text into which the model splits everything it reads and writes (according to Anthropic, in older models one token is on average about 3.5 characters of English text, and models from Claude 4.7 onward split the same text into about 30% more tokens, while the number of tokens also depends on the language). The model turns images into tokens as well. You pay separately for the tokens you send (the prompt, a document, a photo) and for those the model writes in its answer, and the answer is markedly more expensive: Claude Sonnet 5.5 currently costs $2 per million input tokens and $10 per million output tokens (I give prices in dollars, as the providers do).

On top of that come things you don't see in the answer: if the model “thinks” first, all of that thinking is counted as output tokens, even if you don't see it at all or see only a summary of it, and web search at Anthropic and OpenAI costs $10 per thousand searches, plus tokens for the results it finds. An agent that works in many steps sends the whole conversation so far all over again at each one (Claude Code works this way too), so each step usually costs more than the one before.

The cost of analysing documents is a range, not a single amount

My own data shows this best, with one caveat: the processes described below run at my home on local AI models, so I am working out what I would pay if I used the API, at Claude Sonnet 5.5 rates, and the token counts come from my models (a cloud model would count them a little differently), which makes these amounts an order of magnitude and an illustration, not a price list.

The simplest is a process that reads the most important data from my household documents and checks each result exactly once: in a single run over 383 documents, a typical document would cost under a cent, the most expensive about 4.5 cents (more than five times as much), and the whole run about $3.40. The length of the reply in the check alone varies between documents more than twelvefold (42 tokens for one, over 500 for another). A process that reads nutrition values from photos of packaging has an even wider spread in cents, because each product means a different number of photos and a different label: a typical request would cost about 2.4 cents, the most expensive about 10.7 cents, and processing the photos from October's shopping (2,677 requests) about $74 with the Sonnet model and about $148 with the more expensive Claude Opus 5.5.

The pages of one retail chain's promotional leaflets show a different kind of uncertainty: every page carries a similar amount of input data, but the length of the answer ranges from a few tokens to more than four thousand, because on more than a third of the pages (2,523 of 7,048) the model found no food at all, while a page full of products produces a long list. A typical page would cost about 1.3 cents, the most expensive about 5.7 cents, and all the pages from September about $22. When the result is checked and corrected if it is wrong, there is also the number of corrections, which you don't know in advance either (in my documents there is a single check, but allowing further attempts is enough for that cost to start varying too). You can estimate the cost of a whole set quite well, but you learn the cost of a single document only after the fact.

Web search widens the range even further

Gathering information on the web (about a company, a product or a market) is, in my opinion, even harder to price, because the agent itself decides how many times to search, which pages to open and when it knows enough. Every search is a separate fee, every result it finds means more tokens in the following steps, and since the agent sends the whole conversation again at every step, a few extra searches raise the cost of a task not by a little, but markedly.

Images, texts and songs: taste decides how many attempts it takes

In creative work (articles, graphics, videos, music) taste comes in: data from a document is either correct or not, but an image can be correct and still not fit at all, and it is a person who decides whether the first version is enough or only the tenth. On top of that, generated images sometimes have artifacts (small distortions in a random spot) that make a version which looks good at first glance unfit for either a website or a video.

In my image process an automatic judge accepts or rejects every generated image, and depending on the topic one accepted image takes on average from one to more than five attempts (on one of the topics 1,403 attempts produced 594 accepted images), while in another process, with no automatic judge, I deleted 330 of about 2,200 generated images myself (about 15%). If every attempt were generated at the “high” quality of the GPT Image 2 model through OpenAI's API (about 21 cents for a 1024 × 1024 image), a finished image would cost, depending on the topic, from about 21 cents to about $1.10. Songs are similar: sometimes the first go is enough, sometimes it takes three or ten attempts (the first song I published on YouTube I picked from twenty-four versions in three genres). In work like this I know the cost of one attempt in advance, but the number of attempts only shows after the first batch for a given topic.

When API cost can be predicted

I don't want this to sound as if every API cost were a lottery. A task that looks the same every time (the same prompt, data of similar length, an answer of similar length) has a very predictable cost, and it is exactly such a task that one of my real API expenses comes from: a program that every ten minutes, when my home graphics card has nothing to do, asks a local model, and every second time also Claude Sonnet 5.5 through the API, for one deliberately strange idea for an image. In just under three weeks it sent 748 requests through the API, and according to the cost the program works out from the API price list and records with each request, I spent about $27 on it, which is about 3.6 cents per request on average. The most expensive one cost about 4.6 cents, because since October 23 the program has been sending almost the same amount of text every time (the longest request was only about 4% longer than a typical one, while in the first days the requests were shorter and cheaper). For a task like this it is enough to multiply the average cost of a request by the number of requests.

How I would estimate the cost of AI before anything is built

To estimate the cost of AI for a particular company, I would start with a sample of its real data (including the illegible and unusual cases, not only the most textbook ones), run the process on it and work out what a typical item costs and what the most expensive one costs. On that basis I would give a range (this much if the items are mostly typical, and this much if a lot of difficult ones turn up), noting that it is an estimate as of today, because providers change their prices. If the result isn't needed right away, some providers lower the price for batch processing (Anthropic and OpenAI by half, xAI by 20% on selected Grok models). Finally, I would set spending limits and alerts on the API account, so that going over budget ends with a message rather than a surprise on the bill.

A subscription at $20 or $100 is, in my opinion, a good price for a tool for one person, but it is not an AI budget for a company, because that depends on the data, on the number of attempts and on taste, and the most honest way to give it is as a range. If you are wondering what this could look like for you, I have gathered on the homepage a short list of what is worth knowing before we talk.

In short

• A subscription to Claude Code, ChatGPT, Grok or Gemini is designed for a person at the keyboard and has limits, and Anthropic states that it may change them at its discretion.
• Under Anthropic's terms, automation and products for customers are meant to run through an API key, and the other providers also bill the API separately from their subscriptions.
• With the API you pay for the tokens you send and for those the model writes in its answer (which cost more), and the model's thinking and web search cost more than the answer itself shows.
• A task that looks the same every time has a predictable cost: one of my real API expenses is about $27 for 748 requests from the same program.
• I would estimate the cost of AI in a company on a sample of its real data, give the cost of a typical item and of the most expensive one, and set spending limits on the API account.