At API prices, my October in Claude Code would have cost 38 times my subscription

Lately, in conversations with friends and with people who are thinking about bringing AI into their company, I hear something like “I pay $200 a month and I have AI”, as if the price of a subscription to ChatGPT, Codex or Claude Code (an AI coding agent) told you how much artificial intelligence really costs. I pay $200 a month for Claude's Ultra plan myself (a personal preference, not a recommendation), and what I used of it in Claude Code in October would, according to the API price list (the rates at which Anthropic sells programs access to its models, billing every use), have cost about $7,564, which is about 38 times the subscription. It seems to me that providers heavily subsidize plans like this so that users get used to AI and use these tools more and more, but that is my interpretation, not a fact, because the API price list shows what the provider sells tokens for (tokens are the pieces of text into which a model splits what it reads and writes), not what serving the subscription itself costs it, and that cannot be seen from the outside.

What a Claude subscription costs and what limits it has

Anthropic's paid plans are Pro at $20 a month, Max at $100 and Ultra at $200, and each of them includes Claude Code, which draws on the same usage pool as the Claude chat. Pro and Max have two limits: a session limit that resets every five hours, and a weekly one (verified Ultra accounts have only the weekly one), and Anthropic reserves the right to change them at its discretion. Max and Ultra also give a monthly credit for the API platform ($100 or $200), but that is a separate pool: Anthropic points out that it cannot be used in Claude Code or in the Claude apps. Once the limit runs out you can keep working on usage credits, but those are billed at standard API rates and appear on the bill separately, as a charge on top of the subscription. In my opinion a plan like this is designed for a person at the keyboard, not for a program that works for a company on its own. The Claude Code documentation puts it more narrowly: the advertised limits of the Pro, Max and Ultra plans assume “ordinary, individual usage”, and developers who build products or services that use Claude should connect with an API key.

How to convert Claude Code usage into API prices

Claude Code records every session in files on the computer, including how many tokens each of the model's responses used, broken down by type of token (because each type has a different rate in the price list). At my request an AI agent wrote a program that read all such files from October on my laptop, counted each model response only once and priced it according to Anthropic's API price list, and then compared the result with ccusage, a separate tool that works out usage from the same files in its own way and came to $7,563.71 (once one bug in the program was fixed, the two agree). Claude Code itself counts this way, too: the /usage command shows the cost of a session in dollars, worked out on the computer from the number of tokens according to the price list, and Anthropic writes in its documentation that Pro, Max and Ultra subscribers have their usage included in the subscription, so that figure does not matter to them for billing.

The roughly $7,564 is neither a bill nor a saving; it only answers the question of what the same usage would have cost if it were paid for through the API. It is also a lower bound on what I used of the plan in October, because the result covers one laptop and Claude Code alone (without the chat in the Claude app), and the data from the first days of October is incomplete (at startup Claude Code deletes session files older than 30 days).

What my October would have cost at API prices

In October my 261 sessions received 83,805 responses from AI models and used 19.43 billion tokens in total. Almost all of the list-price total, about $7,323, comes from Claude Opus 5.5, and about $232 from the more expensive Claude Fable 5.1.

In each of the three full weeks of October my usage would have cost between $1,313 and $1,925 at list price, against about $46 a week for the subscription, and the most expensive day, October 31, would have cost about $863, which is more than four months of the subscription. That is heavy usage, because I often have several AI agents working at once (62% of the tokens were used by helper agents started by my sessions, and the rest by the sessions I worked in myself), so someone who asks a model a few questions a day will probably use only a fraction of what I do.

Why Claude Code uses so many tokens

Of those 19.43 billion tokens the models wrote only 49.8 million in their responses, and 97% were reads of the conversation from the cache. With every request Claude Code sends the model the whole conversation so far, and it does so again every time the model uses a tool (reads a file or runs a command, for example), so even a one-sentence question in a session that has been open all day uses tokens for the whole conversation. To avoid processing the same history from scratch every time, the API reads it from the cache at a fraction of the normal price of the tokens sent to the model, but with this many tokens such reads still come to about half of the whole list-price total.

An API list price is not the provider's cost

The ratio of about 38 to 1 is measured, but the API price list shows the selling price, and Anthropic does not say what it costs to serve a plan like this. The closest thing to an answer is the S-1 registration statement the company filed with the US Securities and Exchange Commission (SEC): in the first half of 2026 Anthropic had $16.29 billion in revenue, and its cost of revenue, which consists almost entirely of spending on the computing power needed to serve customers, was $2.38 billion, about 15% of that amount. That, however, is an average across all products, with no split between the API and subscriptions, so the document does not show how much Anthropic makes or loses on subscriptions alone. What the same document does show is that in the second quarter of 2026 Anthropic as a whole company had an operating profit of $1.0 billion, at an operating margin of 8%.

What providers and analysts say about subscriptions

What speaks for my interpretation, though, is what the providers themselves write: in July 2025 Anthropic said on X that some of Claude Code's biggest fans were running it continuously in the background, around the clock, and that a few such outlying cases were very costly to support (one user consumed “tens of thousands in model usage” on what was then the $200 plan; the post does not say tens of thousands of what). In the same thread it announced weekly limits for the Pro and Max plans from late August, which it estimated would affect fewer than 5% of subscribers. In the S-1, in turn, Anthropic writes that people who use Claude for their own work and then champion it in their companies have strategic value for Anthropic as a “top-of-funnel engine” for its enterprise business. It also notes that subscribers who approach or use up the usage included in their plan can move to a higher plan or buy more usage, and it calls this “a seamless bridge from subscription to incremental pay-for-intelligence revenue”. In its risk factors it adds that its agent-based products are structurally more compute-intensive than those without agents.

Analysts at SemiAnalysis, who bought these plans themselves and converted them into API prices, wrote on October 5, 2026, that subscription plans are “highly subsidized”, but can make economic sense as a powerful tool for acquiring customers and for marketing. Two days later, in the description of an episode of their podcast, they said that from Anthropic's $200 plan you can pull Opus 5.5 tokens worth $12,000 at API prices (60 times the fee), but “the catch isn't subsidy”, because in their view the API itself is what is overpriced. So within two days the same firm called subscription plans “highly subsidized” and wrote that Anthropic makes a 50-60% gross margin on its own plans, which to me shows clearly that from the outside you can measure the ratio of the fee to the value of the usage at list prices, but not what serving a plan really costs the provider.

ChatGPT and Codex subscriptions look similar

At OpenAI the rules are similar: Codex (an AI coding agent) is included in every ChatGPT plan, the free one too (the Pro plan costs $100, $200 or $500 a month), and ChatGPT plans are billed separately from the API. OpenAI even writes in the Codex documentation that API token prices are separate from subscription usage and should not be used to estimate how many tasks fit into a plan, whereas what I am doing here is pricing usage that has already happened. In January 2025 Sam Altman of OpenAI wrote on X: “insane thing: we are currently losing money on openai pro subscriptions!”, but that post is nearly two years old and was about the Pro plan as it was then.

The price of a subscription is not the price of AI in a company

The price of a subscription tells you little about what AI costs when it works for a company. A program that is meant to run in a company with no person at the keyboard (answering customer requests or reading data from documents, for example) usually uses the API and pays for tokens (sometimes at a discount, such as for batch processing) according to the same price list I used here to price my October. My usage is a developer's work with AI agents, and a different process will use more or fewer tokens, but the price list will be the same. On top of that, how generous a subscription is depends on a decision by the provider, which can change it and has done so before (as with the weekly limits in 2025).

For one person at the keyboard I think a plan like this is well priced, but I would not base a company's automation budget on it; I would use the API price list and a sample of real data. On the homepage I have briefly described how I would estimate the monthly cost of AI in your company and why I would rather give an estimate than promise a fixed amount.

In short

• According to Anthropic's API price list, what I used in Claude Code in October on the $200 Ultra plan would have cost about $7,564, about 38 times the subscription (a lower bound, because the result covers only one laptop).
• With every request Claude Code sends the model the whole conversation so far, which is why reads from the cache made up 97% of my October tokens and about half of the list-price total.
• An API list price is a selling price, not the provider's cost, and Anthropic gives its cost of revenue, which consists almost entirely of spending on computing power to serve customers, only for the company as a whole (about 15% of revenue in the first half of 2026), with no split between the API and subscriptions, so the idea that providers subsidize subscriptions is my interpretation, not a fact.
• Anthropic itself has written that a few cases of continuous Claude Code use are very costly to support, and it calls individual users the top of the funnel for its enterprise business, while within two days analysts at SemiAnalysis called subscription plans “highly subsidized” and wrote that Anthropic makes a 50-60% gross margin on its own plans.
• A program that works for a company usually pays for tokens according to the API price list, and how much fits into a subscription is something the provider can change at its discretion, so I would not base a company's AI budget on the price of a plan.