Pricing guide · billed in dollars of AI usage
Claude Code pricing & usage limits, explained (2026)
Short answer: Claude Code comes with a Claude subscription — Pro at $20/mo ($17/mo if you pay annually), Max 5× at $100/mo, Max 20× at $200/mo. Your subscription buys usage limits rather than a published allowance of tokens — the chunks of text the AI reads and writes. Once those limits are reached you can buy usage credits and carry on at pay-as-you-go prices, the per-token rates Anthropic charges through its API.
Anthropic doesn't publish a quota. Usage runs against a rolling five-hour window with weekly caps behind it, and the Max tiers are described only as 5× and 20× the Pro limits (as of Aug 2026). That vagueness is the defining feature of this platform's pricing, and it's why the numbers on this site are given in dollars of pay-as-you-go equivalent — what the same work would cost billed per token — rather than in anything Anthropic invoices you for.
For the receipts we assume Pro absorbs roughly $40/mo of that work before the weekly caps bite, with Max at 5× and 20× that. That figure is our estimate, not a published number — it is the single biggest assumption on this page, and everything after it inherits the uncertainty.
Claude Code plans
| Plan | Price | What you get |
|---|---|---|
| Pro | $20 per month ($17/mo billed annually) | Claude Code in the terminal — the text-only window where you type commands — plus the Claude apps. Limits are rolling 5-hour windows with a weekly cap on top; no token allowance is published. |
| Max 5× | $100 per month | Five times the Pro limits, same window-and-cap structure (as of Aug 2026). |
| Max 20× | $200 per month | Twenty times the Pro limits. The plan for people who run the agent most of the working day. |
| Pay as you go (the Claude API) | Per token | The metered path: open a pay-as-you-go account with Anthropic (an API key) or switch on usage credits once the subscription limit is hit. Opus 5 is $5 per million tokens read and $25 per million written; Sonnet 5 is temporarily $2/$10 through Aug 31, 2026, then $3/$15; Haiku 4.5 is $1/$5. |
Hit the limit and you can wait for it to reset, upgrade to Max, switch on usage credits at pay-as-you-go prices, or open a pay-as-you-go account with Anthropic (an API key). The metered options turn a flat subscription into a variable bill, so set a limit before a long session.
How Claude Code usage limits actually work
There is no unit to count. Claude Code doesn't meter credits and doesn't show you a balance ticking down — it runs until you hit a limit, and that limit is a rolling five-hour window, with a weekly cap sitting behind it for heavy weeks.
The plans are sold as multiples rather than amounts. Max 5× and Max 20× buy five and twenty times the Pro limits, and Anthropic doesn't publish the underlying allowance for any tier, so nobody outside the company can turn a plan into features or hours with confidence (as of Aug 2026).
So we model it, and we say so. Our assumption: Pro covers about $40/mo of build work priced at pay-as-you-go rates before the weekly caps bite, and the Max tiers cover 5× and 20× that. It is an estimate worked out from those published per-token prices, not an allowance Anthropic has confirmed; Sonnet 5's temporary introductory rate also means the equivalent can change after Aug 31, 2026.
Every Claude Code figure on this site is therefore a pay-as-you-go equivalent: what the same work would cost billed per token. On a subscription that's a capacity number, not an invoice. Your invoice is a flat $20, $100 or $200, and the thing you're really spending is calendar time.
What features cost on Claude Code
Base cost per feature in dollars of AI usage, before the iteration tax (the extra rounds nobody escapes) and the debug surcharge (fixing what breaks). These are the ranges our estimator adds up.
| Line item | Cost in dollars of AI usage |
|---|---|
| Core app — simple | $6–$14 |
| Core app — standard | $10–$22 |
| Core app — complex | $18–$36 |
| User accounts & login | $5–$10 |
| Payments / subscriptions | $7–$14 |
| Admin panel | $6–$12 |
| File & image uploads | $3–$7 |
| AI features | $6–$13 |
| Connections to other services | $5–$11 |
| Multi-user roles | $4–$8 |
Multiply by your prompting style (×1.2 to ×8), then add 40% for debugging. See the full formula →
The gotcha
There is no published allowance to count down: Claude uses rolling 5-hour windows plus weekly limits. When a paid plan hits its limit you can keep going by buying usage credits, charged at direct pay-as-you-go AI prices. The capacity shown here is our estimate, not a number Anthropic publishes.
The pricing is simple; the budgeting isn't. With no credit counter there's nothing to watch tick down, so the first signal that you've spent the week's capacity is the agent telling you to come back later — usually mid-refactor (halfway through restructuring the code), and usually on the evening you had free.
All three ways out cost something: wait out the rolling window, jump to Max at $100–$200/mo, or fall back to pay-as-you-go prices and let the bill float. Our months-at-plan arithmetic assumes you wait, which is why a Claude Code receipt stretches the calendar further than it stretches the money.
“Cheapest per feature” also deserves its asterisk — it isn't ours to claim. Against Replit and v0 it holds, because nothing sits between you and the model taking a cut. Against Cursor + Claude it doesn't, and against Lovable credits and Bolt tokens at the conversion rates we publish it doesn't either: our receipts put all three lower in dollars on most builds, and we let the arithmetic say so rather than the marketing.
The other half of the gotcha is the skill floor. Claude Code will happily produce a week of work you don't understand, and the cheapness is real only if you can tell a good change from a plausible one. If you can't, the cheap-looking platform quietly becomes the most expensive one.
Who Claude Code is for
People who can read the code it writes, and are willing to. It is cheap per unit of work and carries a flat, predictable invoice — though Cursor + Claude runs cheaper still on our conversions, since Cursor's $20 buys usage rather than sitting on top of it — and it carries the steepest learning curve of the six: an agent in the terminal with no live preview, no visual editor and nothing to catch a bad change except your own review. On our conversions the credit- and token-priced builders still land lower in dollar terms; what you buy here is control, a flat invoice, and no platform mark-up.
See real numbers
Printed receipts for three common builds on Claude Code:
Get your build’s receipt
Pick what you’re building, pick how you prompt, and see the bill in Claude Code usage limits before you spend any.
A plain link to the platform. Nothing on this page changes which one we call cheapest.