Claude Code pricing: what it really costs, with our own test

Claude Code is Anthropic's AI helper for programmers. It works in your terminal: you tell it what to build or fix, and it reads your files, writes code and runs tests. It is not free. You need a paid Claude plan, which starts at $17 a month, or you pay for each use. We ran the same small coding tasks with three Claude models. Then we worked out what one task really costs: from about 14 cents with the smallest model to about 54 cents with the biggest.

Is Claude Code free?

No. The free Claude plan does not include Claude Code. Every paid plan does, Anthropic says on its pricing page.

There is one catch. Claude Code shares its limit with the rest of your plan. So if you chat with Claude a lot on your phone, you have less left for coding, and the other way around.

You can also skip the plan and pay for each use with credits in a developer account. Then you pay API prices, which you can find further down. An API is the connection that programs use to talk to Claude.

The plans for one person

These are the prices in the US, before tax, on 6 October 2026. In Europe we saw prices in euros.

• Pro: $20 a month. If you pay for a whole year at once ($200), it works out at about $17 a month.
• Max 5x: $100 a month, for five times as much use as Pro.
• Max 20x: $200 a month, for twenty times as much use as Pro.

Max is only sold per month, the help page on Max says. It also gives you new features and models earlier. Prices in the phone apps can be different.

The plans for teams

Teams of 2 to 150 people can choose a Team plan. You pay per person, called a seat:

• Standard seat: $25 a month, or $20 a month with a yearly plan. A bit more use than Pro.
• Premium seat: $125 a month, or $100 with a yearly plan. Five times a Standard seat.

You can mix both kinds in one team. Big companies can choose Enterprise: $20 per person a month, plus the cost of everything they use, at API prices.

How the limits work

Anthropic does not say how many messages or tasks you get. It depends on how long your chats are, which model you pick and what you ask, the pricing page explains. What it does say:

• Your limit resets in a window of five hours that keeps moving.
• Paid plans also have a weekly limit on top.
• Pro gives at least five times as much use per window as the free plan.

When you hit the limit, you can wait, move to a bigger plan, or turn on extra credits that you pay at API prices. You can see how much you have left under Settings and then Usage.

Paying per use: the API prices

If you pay per use, you pay per million tokens. A million tokens is about 750,000 words: several thick books. Here are the prices for the newest models, from Anthropic's price list:

• Haiku 4.5: $1 to read, $5 to write. The smallest and cheapest.
• Sonnet 5.5: $2 to read, $10 to write.
• Opus 5.5: $4 to read, $20 to write.
• Fable 5.1: $10 to read, $50 to write. The biggest.

There is one more thing to know. Claude Code keeps the whole conversation and your code in a kind of short-term memory, called a cache. Reading from the cache is much cheaper than reading new text: $0.20 per million tokens for Sonnet instead of $2. Saving to the cache costs extra. Claude Code saves for one hour at a time, and that costs twice the normal price for reading new text.

Our test: what one coding task costs

We gave Claude Code the same three small tasks as in our Codex vs Claude Code test. One was to fix a bug, one to build an invoice calculation from a description, and one to add due dates to a small to-do app. Each model did each task twice. We checked the work with secret tests: tests we wrote ourselves and only ran after the AI had finished.

Claude Code reports how many tokens it used. We multiplied those by the list prices. That gives the cost if you pay per use:

• Haiku 4.5: about 14 cents per task. It passed 80 of 88 secret tests.
• Sonnet 5.5: about 28 cents per task. It passed 86 of 88.
• Opus 5.5: about 54 cents per task. It passed 86 of 88.

Sonnet and Opus passed the same secret tests. On the bug task, both missed just one test in each try: the same test every model in our earlier tests missed. Haiku is half the price of Sonnet, but it missed more: on the bug task it passed 21 and 19 of 24. So the cheapest model was not the best deal for that task. For the bigger tasks it did fine.

Want to see the test? One secret test checks that "1h banana" is refused, because there is junk after the time. Another checks that an invoice line of 1.005 kilos is allowed. Our test page shows all three jobs, the secret tests and the exact words we sent. You can also download them and run them on any AI yourself.

Where the money goes

The biggest cost was not the code Claude wrote. It was the context: all the files and messages Claude Code keeps reading. With Sonnet 5.5, 59% of the bill was for saving that context to the cache, and 24% for reading it back. Writing the answers and the code was only 17%. About 89% of everything it read came from the cheap cache; without it, the same task would cost more than twice as much.

That is why long sessions cost more than short ones. Every step, Claude Code reads the whole conversation again. Our tasks were small. A big project with many files and a long chat uses far more tokens per step.

Which one should you choose?

At API prices, $20 buys about 71 tasks like ours with Sonnet 5.5, about 37 with Opus 5.5 and about 142 with Haiku 4.5. That is not the number of tasks in a Pro plan: Anthropic does not publish that number, and real work is usually bigger than our tests.

Our advice:

• Trying it out, or coding a few times a week: Pro is the simplest choice. You know what you pay.
• Coding every day, and you keep hitting the limit: look at Max, or pay per use if your work comes in bursts.
• Want to save money: a smaller model like Haiku costs about half as much. But in our test it made more mistakes, so check its work.

Want to compare with others? In our DeepSeek V4 test similar tasks cost a few cents each. That test used a simpler setup without Claude Code, and DeepSeek has other trade-offs.

How we checked

On 6 October 2026 we read Anthropic's pricing page, its API price list and the help pages on Pro and Max. The screenshots are our own captures of those pages, cropped only.

We ran Claude Code on 5 and 6 October 2026, with default settings and the owner's own subscription, so we did not pay per use. Costs are worked out from Claude Code's token counts and the list prices. We do not show how long each task took: our computer was busy during part of the test, so the times were not fair. One Haiku try got stuck for 40 minutes without an answer, so we ran it again. For two Opus tries the first check of the secret tests ran out of time, so we ran those checks again by hand. The charts were made by us. None of the images was made by AI.

This was a small test with three short tasks. It says nothing about big projects, and prices and limits change. Check Anthropic's page before you choose a plan.