How much does Claude Code cost?

On a subscription it is flat: $20/month on Pro, $100 on Max 5x, $200 on Max 20x, and the bill does not move with how much you run. On the API you pay per token, and what that comes to depends far more on how you feed the model than on which model you pick. One real job, priced from its own measured token counts, came to $452 sent as raw HTML on the standard API and $47 sent as extracted text on the Batch API.

Is the $20 Claude Code plan worth it?

It is not a cost question, because the bill is the same whatever you run. What you are buying on a higher tier is not being stopped mid-task by the rolling 5-hour and weekly usage limits. Upgrade when the limit interrupts you repeatedly on ordinary days, not because a calculator says so.

Is the Claude API cheaper than a subscription?

It depends entirely on volume, and the two meters are not comparable. A subscription is capped and rate-limited with no per-token record; the API is uncapped and metered. Price your actual job on the API first, then compare it to the flat plan fee.

Claude Code pricing

Two meters, and only one of them has arithmetic

On a subscription the bill is flat. It does not move with how much you run, so there is nothing to calculate. On the API you pay per token, and there the same job prices out at $11.17 or $109 depending on two decisions. This page answers the first honestly, then prices the second.

Subscription Flat $20 / $100 / $200 a month. What runs out is time, not money. API Per token Metered and uncapped. This is the one worth calculating.

The plans

What each tier costs

Plan	Monthly	Billed annually
Pro	$20	$17	Includes Claude Code.
Max 5x	$100	—	claude.com/pricing shows "From $100 per month".
Max 20x	$200	—	claude.com/pricing groups both Max tiers under "From $100 per month". $200 is the figure on an actual Max 20x bill, before state tax.
Team (standard seat)	$25 /seat	$20 /seat	Includes Claude Code and Cowork.
Team (premium seat)	$125 /seat	$100 /seat	Per seat.

Why there is no plan calculator on this page

Every competing page runs arithmetic here — dollars per message, cost per Pro-equivalent unit, break-even tables. The arithmetic is fiction. A subscription is capped and keeps no per-token record, so there is no usage curve to plot and no break-even to find. The bill is $200 whether you run one session or five agents across three worktrees.

What you actually buy on a higher tier is not being stopped mid-task. There are two ceilings, a rolling 5-hour window and a weekly one, and they are shared across every Claude surface — chat on claude.ai spends the same allowance as your terminal. Upgrade when the limit interrupts you repeatedly on ordinary days. That is the whole decision, and no number on this page will make it for you.

Limits per Anthropic's Pro and Max plan documentation. Plan prices from claude.com/pricing, read 2026-10-21.

The flat bill has a door out of it, and it opens onto the other meter

When a paid plan hits its limit you are not simply stopped any more. Pro, Max 5x and Max 20x can switch to usage credits and keep going, and Anthropic bills those at standard API rates, charged separately from the subscription. Pre-purchased usage bundles discount that overflow.

So the two meters on this page are not really alternatives. The subscription is a flat fee with a ceiling, and above the ceiling you are on the per-token meter whether you planned to be or not. That is the number the calculator below gives you, and it is worth knowing before you hit the ceiling rather than after.

Usage credits on paid plans and usage bundles, Anthropic Help Center, read 2026-10-21.

Paying per token instead? Here is what a job really costs.

I read page one for this term on 2026-08-22. All nine results reprint the same per-token rates. Rates are not costs. A rate becomes a cost only when you multiply it by how many tokens your work actually is, and none of the nine puts a number on that, because doing so means having run something and counted. Below is a calculator built on counts we took with Anthropic's own count_tokens endpoint over 2,399 real pages.

The calculator

It asks how much work you have, not how many tokens

There is no token field on purpose. Asking you for a token count is asking you for the answer you came here to get.

1 What are you doing?

2 How much of it?

pages

A web page, raw HTML — 79,769 tokens in, 3,400 out, each. measured output inferred

3 How are you running it?

Model Most production workloads. The default unless you have a reason. Send extracted text, not raw HTML Same checks. 7.5x fewer tokens, because the checks that need markup are ones a parser answers better and for free. Batch API — 50% off, up to 24h wait Results come back in any order, no streaming, 256 MB per batch, and unfinished requests expire rather than queue. Not available for this job — an interactive session is not batchable. Cache hits 0% A cache read costs a tenth of a fresh input token. This is the lever on long agentic runs, where the same context is resent every turn.

This job, as configured

$109

Includes 12.5% for retries and failed fetches. Every measured figure assumes one clean pass, which is not a thing that happens.

The spread on this job

$11.17 to $109

Same job, same checks, same model. The gap above is what you send and when you run it. Changing the model moves the whole range, which is why it is held fixed here.

The one change that moves this most

Send extracted text instead of raw HTML. Same checks, 7.5x fewer tokens.

Worth $86.53 on this job.

Before any of this: the cheapest call is the one you never make

On the job these numbers come from, roughly two-thirds of the checklist needed a lookup, not judgment — status codes, link inventories, whether the HTML parses, whether a script fired. Those cost $0 in API spend because a script answers them exactly, for free, in milliseconds. Splitting the list took an afternoon and took more off the bill than every other decision here combined.

So the honest first move is not choosing a cheaper model. It is deleting the calls that never needed a model.

The part the other nine leave out

A cheaper model is not proportionally cheaper

Claude 4.7 and later, which includes Sonnet 5, Opus 5 and Fable 5, use a newer tokenizer that produces about 30% more tokens for the same text than Sonnet 4.6 and earlier. Haiku 4.5 is on the older one.

Two consequences, neither of which appeared on any of the nine pages I read. First, a token count measured on one model does not transfer to another, so a per-page figure you took on Sonnet 5 will overstate the same page on Haiku 4.5. Second, the gap between models is wider than the rate card alone: Haiku 4.5 is half Sonnet 5's input rate on paper, and cheaper again once the same text tokenises shorter. This calculator applies the correction where it belongs and nowhere else. A figure Anthropic published for Haiku, or a bytes-to-tokens rule of thumb, is left alone.

Anthropic's wording is "approximately 30%… The exact increase depends on the content and workload shape." Treat it as a correction with a margin, not a constant — and re-measure your own corpus rather than inheriting ours. Source, read 2026-10-21.

Line items that are not in any rate table

Charge	Rate
Web search	$10 per 1,000 searches	Plus standard token cost for the content retrieved.
Web fetch	— no additional charge	You still pay tokens for whatever it pulls in.
Managed Agents session runtime	$0.08 per session-hour	Accrues only while status is `running`; idle time is free.
Code execution	$0.05 per container-hour after 1,550 free hours/month	Free when used alongside web search or web fetch.
US data residency	1.1x on every token category	Applies when `inference_geo` is pinned to "us" on Claude 4.6+.

Method, and what this does not cover

Where every number here comes from

Unit of work	Input	Output	Source
A web page, raw HTMLThe whole document — classes, inline SVG, analytics snippets	79,769	3,400	measured 79,769 tokens on an English marketing page, 77,028 on a translated one, across 2,399 URLs. Mean 196,962 bytes of raw HTML per page. Output is the QA report the run produced, 3,400 tokens per page.
A web page, extracted textVisible copy plus structured facts — the markup thrown away	10,600	1,850	measured output inferred Input counted: 10,600 tokens for the same pages as extracted text and structured facts, 7.5x less than raw HTML for an identical checklist. Output is INFERRED, not counted: the run reconciles to its published $95 standard / $47 batch only at ~1,850 output tokens a page, which is what moving the link inventory out of the model job would do. At the 3,400 figure the raw-HTML run produced, the same arithmetic gives $132 and $66.
A web page, visible copy onlyJust the words a reader sees, nothing structured	8,290	1,850	measured 8,290 tokens of visible text on an English page. Output shrinks here too: strip the link inventory out of the model's job and the report roughly halves.
A fetched web page (10 kB)Anthropic's own estimate for an average page	2,500	800	Anthropic Anthropic's web fetch tool documentation: "Average web page (10 kB): ~2,500 tokens".
A documentation page (100 kB)A long reference page, fetched whole	25,000	1,500	Anthropic Anthropic's web fetch tool documentation: "Large documentation page (100 kB): ~25,000 tokens".
A research paper PDF (500 kB)A full academic paper	125,000	2,000	Anthropic Anthropic's web fetch tool documentation: "Research paper PDF (500 kB): ~125,000 tokens".
A support conversationOne ticket, end to end	3,700	—	Anthropic Anthropic's pricing page worked example: "Average ~3,700 tokens per conversation", costing ~$37 per 10,000 tickets on Haiku 4.5. The 3,700 is the whole conversation, so output is folded in rather than counted separately.
An hour of agent codingAnthropic's own worked example for a managed session	50,000	15,000	Anthropic Anthropic's Managed Agents worked example: a one-hour coding session on Opus 5 consuming 50,000 input and 15,000 output tokens. Note this EXCLUDES the $0.08 per session-hour runtime charge, which the calculator adds separately.

What would make this wrong for you

The anchors are one corpus. They were measured on a single multilingual marketing site — 2,399 pages, mean 197 kB of raw HTML each. A JavaScript-heavy app, a codebase, or a set of scanned PDFs will not tokenise anything like it. Treat them as a starting point you replace, not a constant.
Output tokens are the soft number. Input we counted exactly. Output depends on what you ask for, and a verbose report format can double it without changing the job.
Long agentic loops are a different shape of bill. There the driver is conversation history resent every turn, so the lever is caching rather than extraction. That case is modelled here only through the cache slider.
Prices move. These were checked on 2026-10-21. The figures this page is built on were themselves 1.5x too high within a day of being measured, because a scheduled rate increase was cancelled. Nothing about a stale price looks stale.

Rates verified 2026-10-21 against Anthropic's pricing page and claude.com/pricing. A checker runs on every deploy and blocks it if any rate here has moved.

The full write-up, with the token counts and the reconciliation: the same job priced out at $452 and $47. On the plan side: Pro vs Max 20x is not a cost decision.