Claude Pricing: Subscription or API?

Short answer

• What it costs: from Switzerland, OpenAI shows CHF 20 a month for ChatGPT Plus and CHF 83 for the smallest Pro tier, which costs USD 100 in the US. Claude Pro costs USD 20 a month, Max 5x USD 100 and Max 20x USD 200. A Standard seat on a company plan costs CHF 19 per person a month on ChatGPT Business and USD 25 on Claude Team, less when paid annually. Anthropic's prices exclude tax.
• Plan or API: the API is the access a program uses to talk to a language model directly, without a chat window. If you do a lot of AI work yourself at your computer, an individual plan is almost always cheaper than the API, above all for agentic use, where a model works through many steps on its own.
• Business use: each person needs an account of their own. If you work alone, you can use an individual ChatGPT plan such as Plus or Pro for business. By their wording, the consumer terms Anthropic shows in Switzerland rule out business use of Claude Free, Pro and Max, so if you work alone, your only option on Claude is the API, which you use with an API key (a kind of password for programs). For businesses, Anthropic offers the Team plan for two or more people.
• What a plan is worth at API prices: on 6 October 2026, the research firm SemiAnalysis calculated how much API value a plan gives if you use its full monthly limit for agentic tasks. An individual Claude plan comes to about 57 to 59 times its price with the model Opus 5.5, an individual ChatGPT plan to about 10 times with GPT-6.1 Sol. The company plans Claude Team and ChatGPT Business are not part of this calculation. The two figures cannot be compared directly, because SemiAnalysis prices each vendor with its own model. For Claude Max 20x at the list prices of 29 October 2026, the value ranges from USD 2,485 a month with Fable 5.1, another Anthropic model, to USD 11,726 with Opus 5.5.
• Usage windows: Claude limits usage with a five-hour window and a weekly limit, and you hit the limit as soon as either one is used up. On ChatGPT Plus, the same two windows apply only in the coding tool Codex and in ChatGPT Work, its agent mode. ChatGPT Pro currently has only a weekly limit there.
• API credit: every month, Claude Max 5x comes with USD 100 of API credit, Max 20x with USD 200 and the Team plan with up to USD 500 for all seats together. The credit expires each month and does not roll over.
• A plan for customers: the consumer terms of both vendors forbid making your account available to anyone else. A product for customers runs on the API or, in a few partner apps, on the customers' own ChatGPT plan.

A USD 20 plan, USD 215 of API usage in three days

On 9 October 2026, the X user @coldcat_ai worked out that three days with Claude Pro, a plan at USD 20 a month, would have cost him USD 215 at API list prices for Opus 5.5. He had used 66% of his weekly limit in those three days, working in Claude Code. A token counter came to 678 million tokens. Tokens are the pieces of text, mostly parts of words, that language models are billed by.

He explains why in the same post. "99% of claude's tokens are cache reads. the whole session gets re-read every turn." Almost all of it was text the model read again from its cache at every step, at a fraction of the normal API price. The figures from @coldcat_ai are one user's measurement, not an average. How much this re-reading counts against a plan's limit is up to the vendor.

The USD 215 and the USD 20 a month come from two ways of billing. The plan gives you an allowance in time windows. The API charges for every call, meaning every request a program sends.

In the comments under a Tages-Anzeiger article of 15 August 2026 about AI costs at companies, a reader put it this way (in German, my translation):

"You still pay for access through the web GUI in the monthly plan. But if you connect development environments directly through the API, you pay in what are called tokens."

The web GUI is the chat window in the browser, and development environments are the programs developers write code in. Coding tools such as Claude Code and Codex also run on a plan, though. You pay in tokens when a tool is connected through an API key.

What a plan is worth at API prices

The SemiAnalysis calculation

In an X thread on 6 October 2026, the research firm SemiAnalysis worked out what the full monthly limit of common plans would cost at API list prices. It assumes agentic use, with Anthropic's model Opus 5.5 for Claude and OpenAI's GPT-6.1 Sol for ChatGPT. Cache reads make up 96.6% of the tokens. The calculation starts from the plans' US dollar prices, so every amount in this section is in dollars. From Switzerland, OpenAI shows its plan prices in francs and Anthropic in dollars.

Source: SemiAnalysis on X, part 3 of 6, 6 October 2026 (Swiss time), values read from the chart. Pro 200 after its allowance was cut shortly before the thread (see below). Pro 500 is not in the chart, and neither are the company plans Claude Team and ChatGPT Business. US plan prices from claude.com/pricing, Claude's help page on Max, OpenAI's help page on the Pro tiers and, for ChatGPT Plus, OpenAI's Codex pricing page, read on 29 October 2026.

In this calculation, the Claude plans come to about 58 times their price and the ChatGPT plans to about 10 times. Within one vendor, the ratio stays almost the same. Max 5x costs five times as much as Pro and brings just under five times as much API value. Max 20x costs ten times as much and brings about ten times as much over the month. The name suggests twenty times. Anthropic does promise twenty times the usage of Pro on Max 20x and five times on Max 5x, but per five-hour window. Over the month, the weekly limit sets the ceiling, and Anthropic gives no factor for it.

SemiAnalysis prices each vendor with its own model, and Opus 5.5 costs twice as much per token as GPT-6.1 Sol. So the table does not compare the two vendors directly. Part of the gap between 58 times and 10 times comes from the model's higher price. Even after taking that price difference out, though, the Claude plan gives almost three times as many tokens per dollar. According to the fifth part of the thread, the gap was even larger with Sonnet 5.5, a cheaper Anthropic model, at early October prices. The calculation does not show whether you also get more done.

Why the same figure varies about fivefold by model

In the first part of the same thread, SemiAnalysis qualifies its own table. "It doesn't make sense to say a subscription plan is worth $ X in isolation," it writes. For Max 20x and agentic use, at the early October prices it used, the range across Anthropic's models runs from USD 2,485 with Fable 5.1 to USD 12,529 with Sonnet 5.5.

These are the list prices per million tokens, each for input (what the model reads), output (what it writes) and cache reads:

• Opus 5.5: USD 4, 20 and 0.20
• Sonnet 5.5: USD 2, 10 and 0.10
• Fable 5.1: USD 10, 50 and 0.25
• GPT-6.1 Sol: USD 2, 10 and 0.10

Sources: Anthropic's prices and OpenAI's prices, read on 29 October 2026.

According to SemiAnalysis, each model uses up the plan's allowance at a different speed, but not in proportion to its API price. Max 20x reaches its highest value with Sonnet 5.5 in this calculation. SemiAnalysis used the list prices of early October, when a cache read on Sonnet 5.5 still cost USD 0.20, the same as on Opus 5.5. It now costs USD 0.10. The prices of Opus 5.5 and Fable 5.1 have not changed. The allowance stays the same, but at that price the same tokens on Sonnet 5.5 would come to less than Max 20x reaches with Opus 5.5.

Fable 5.1 costs the most per token of these four models. Even so, Max 20x reaches its lowest value with Fable, because the plan gives so many fewer tokens that the higher price does not make up for it. Part of that comes from SemiAnalysis counting only half the limit for Fable, because on Max at most half of the weekly limit may go to Fable.

The vendor sets the value

In the comments under Tages-Anzeiger and 20 Minuten articles about AI costs, readers keep raising the suspicion that vendors lure people in with cheap plans and cash in later.

SemiAnalysis writes in the fourth part that OpenAI halved the API-equivalent value of Pro 200 in the week before the thread. OpenAI itself gives no factor. Its help page on the Pro tiers only says that new subscriptions get a "lower usage allowance" and that existing ones kept the previous allowance until 29 October 2026.

On its Codex pricing page, OpenAI advises against this kind of arithmetic. The page says "API token prices are separate from subscription usage" and that they should not be used to estimate the tasks a plan includes.

How the five-hour window and the weekly limit work

Two limits at once

Claude counts your usage in two windows, one of five hours and one of a week. OpenAI uses the same two windows on Plus, but only in its coding tool Codex and in ChatGPT Work, the agent mode of ChatGPT. Ordinary chat in ChatGPT has rules of its own. Where both windows apply, OpenAI writes that "you need allowance remaining in both". In Codex and ChatGPT Work, a new five-hour window starts with the first message after the previous one ends (OpenAI help).

Anthropic calls the five-hour window the session limit. According to Claude's help page on Max, it resets "every five hours", and the weekly limit resets at the same time each week. Both count at once on Claude too. I found nothing in Anthropic's help pages on when the five-hour window starts on Claude.

"Then you say 'do it' a couple of times and the limit is reached," a reader wrote in the comments under a 20 Minuten article (in German, my translation). With agentic use, this can happen fast, because every step rereads the whole history so far, and that history grows with every step.

Where ChatGPT and Claude differ

ChatGPT Pro 100, 200 and 500 currently have no five-hour limit in Codex and ChatGPT Work, only a weekly limit. Claude's chat and Claude Code share one allowance.

My overview of ChatGPT usage limits by plan shows how much each ChatGPT plan allows per window. For Claude, the overview of Claude usage limits by plan shows the same as factors, because Anthropic publishes no message counts.

What happens at the limit

Anthropic sells usage credits for Claude, billed at API prices. OpenAI sells credits for extra usage. They are explicitly "not API credits" and stay valid for 12 months. OpenAI also offers eligible Plus and Pro accounts a paid instant reset of the weekly limit in Codex and ChatGPT Work.

I recommend checking the usage display before a longer task to see how much of the five-hour window and the weekly limit is left. If a window is nearly used up, finish what is already running as long as it still fits, and start anything new only after the reset.

When the API is the right choice

As soon as other people use the result

Anthropic's consumer terms say, "You may not ... make your Account available to anyone else." OpenAI's terms for Europe, which also apply in Switzerland, carry almost the same sentence, "You may not share your account credentials or make your account available to anyone else." For Pro, OpenAI adds that "reselling access" is "prohibited". Anthropic's consumer terms also allow automated access ("through automated or non-human means") only with an API key or where Anthropic explicitly permits it.

So every employee needs an account of their own. On ChatGPT, that can be their own Plus or Pro plan. OpenAI's terms for Europe include a separate addendum for business use, which among other things limits OpenAI's liability. On Claude, the consumer terms Anthropic shows in Switzerland exclude business use of Free, Pro and Max ("not use our Services for any commercial or business purposes"). At the same time, Anthropic's help page on VAT explains how to take out an individual plan "as a business". According to Anthropic, the consumer terms apply to a plan taken out that way as well. Until Anthropic explains how that fits the exclusion of business use, I go by the wording of the terms. Anthropic offers the Team and Enterprise plans for business use, and on these, as on ChatGPT Business, the business terms apply instead of the consumer terms.

The chat on my website runs on the API with an API key of its own, not on a plan login, because other people use it. For that API account, I switched off auto-reload (the automatic top-up of credit) and set a fixed spend limit. Once its balance is used up, requests with this key fail instead of charging a card. When that happens, the chat falls back on a reserve, so it does not go down.

For a product that other people use, one exception to the API is "Sign in with ChatGPT". With it, customers who have a ChatGPT plan (Go, Plus or Pro) use their own plan in participating apps. OpenAI describes them as "a limited set of partners".

As soon as customer data runs through the API or through a plan, for example when staff paste customer emails into the chat, the data protection rules apply. I describe the Swiss rules in ChatGPT and Data Protection in Switzerland.

What a seat on a company plan costs

Sources: ChatGPT's pricing page for businesses, opened from Switzerland on 29 October 2026, and Claude's pricing page, Claude's help page on the Team plan and Claude's help page on the Enterprise plan, read on 29 October 2026. Anthropic shows its prices without tax, and OpenAI says nothing about VAT on its franc prices.

A Standard seat on Claude's Team plan gives 1.25 times the usage of Pro per five-hour window, and a Premium seat 6.25 times, slightly more than Max 5x. Both include Claude Code and also have a weekly limit. OpenAI says ChatGPT Business Premium gives five times the usage of Standard in Codex and ChatGPT Work and has no five-hour limit there. On Claude's Enterprise plan, the seat fee covers only access, and all usage is billed at API prices on top.

What a call costs: a worked example

My example is a customer chat that answers 1,000 requests, each with 3,000 tokens of input and 500 tokens of output, at the list prices of 29 October 2026. With Sonnet 5.5 or GPT-6.1 Sol, the input costs USD 6 and the output USD 5, together about USD 11, a little over 1 US cent per request. The Batch API collects requests and works through them later. For the chat, it stays at USD 11, because a chat cannot wait. A nightly analysis of the same size can wait and costs about USD 5.50 through Batch.

Occasional questions would often cost less on the API than a USD 20 plan. Without a chat window, though, the API needs a program that talks to it, and someone has to build or set up that program. For everyday work at your computer, a plan or a seat on a company plan is therefore usually the simpler route, even with little use, provided the terms let you use a plan.

The most common complaint under articles about AI costs at companies was that costs keep rising and nobody can see what the tokens are spent on. On the API, you can work out the price of each call from its tokens. A plan only shows you how much of the limit is left.

When such calls run on a workflow platform, you pay for the platform as well. I compared n8n's pricing with two of my own scheduled jobs, which run without a workflow platform, in n8n Pricing: What It Costs and When It Pays Off.

When a job runs without you

A program that reads receipts at night or pre-sorts requests needs nobody at the computer. On the API, you can work out its cost from the fixed list prices before the job runs. If it does not have to be finished at once, it can run through Batch at half the price.

Under Anthropic's terms, such a job on Claude needs an API key or explicit permission. Anthropic gives that permission for the Agent SDK, a toolkit developers use to build their own AI agents. A job that only you use and that is not for business may currently run on your plan with it, and it then uses up your five-hour window and your weekly limit. For a job that has to run reliably, I still recommend the API. On the plan, the job stops as soon as your own work has used up the limit, unless you have switched on usage credits.

The API credit in Max and Team

Claude Max 5x includes USD 100 of API credit every month on top of the plan's allowance, Max 20x USD 200, and the Team plan up to USD 500 for all seats together. According to Anthropic's documentation, the credit covers the Claude API and the Agent SDK when it runs with an API key instead of the plan login (read on 29 October 2026). It covers neither interactive work in Claude Code nor extra usage in the apps beyond the limit. It expires at the end of the billing cycle, or monthly on an annual plan, and does not roll over.

According to the documentation, if the credit is used up and your API account has no other balance, API requests stop until the next monthly credit arrives. The documentation adds, "Usage is never charged to your Claude plan." Pro gets no API credit. OpenAI bills its API separately from the ChatGPT plans.

A plan for your own work, the API for the product

If there are at least two of you, start your own work on the smallest company tier, ChatGPT Business Standard or Claude Team Standard. Choose the next tier by how often you hit the limits of the one you have.

On your own, take an individual plan such as Plus or Pro on ChatGPT. For business use on Claude, going by the terms above, it is the API. A Claude Pro plan at USD 20, fully used for agentic tasks on Opus 5.5, corresponds to about USD 1,178 of API costs a month, according to SemiAnalysis. On Sonnet 5.5, the same tokens cost half that on the API, about USD 590. You then work in a program that runs on your API key, for example Claude Code or a chat program you set up once. An individual ChatGPT plan is simpler and, used heavily, cheaper.

If you often work in Codex for hours at a stretch, look at ChatGPT Pro and Business Premium, which have no five-hour limit there. For customers, for internal tools that staff use without an AI account of their own, and for programs that have to run reliably without you, I recommend the API with a fixed spend limit.

According to the SemiAnalysis values, an individual plan used heavily for agentic tasks is many times cheaper than the API, and that direction is enough for me to choose between the two. For the company plans, a factor per five-hour window does not tell you what you get in a month. On Max 20x, twenty times the usage of Pro per window came to about ten times Pro over the month in the calculation above. I would not plan a budget on any of these values, because the vendor can change a plan's allowance.

If your own work keeps hitting the limit, I would look at the way you work first, and only then at a larger plan. I think shorter conversations with one clear task do more than moving up a tier, because each step then rereads a shorter history.

Sort your AI tasks into two lists

For one week, keep two lists. The first is for the AI tasks you do yourself at your computer. The second is for the tasks in which customers use an AI tool you set up, staff without an AI account of their own use such a tool, or a job runs without you. The choice between a plan and the API rests on those two lists and the terms above. The AI consulting page shows how I assess tasks like these. You can go through the second list with me in a first call.

Frequently asked questions