Claude Pricing: What Each Plan Really Costs

Claude pricing looks simple on the marketing page. It stops looking simple the moment you try to budget for it.

Anthropic sells five tiers. Three carry a price you can read in two seconds. Two of them bill you for something invisible: tokens.

Here's the gap. The cheapest paid plan is $17 a month. The most expensive plan on the same page has no ceiling at all, because it charges a seat fee and then meters every single thing you type on top of it.

That's not a pricing tier. That's a meter with your name on it.

Which plan is right for you depends on one number Anthropic's pricing page never asks: how much you actually use Claude inside a five-hour sitting.

Here's every current price, what each tier actually includes, where the usage limits bite, and the mistake that quietly turns a $20 plan into a very different bill three months later.

What every Claude plan costs right now

Anthropic's own pricing page lists five tiers this week. Every price excludes tax and is, in Anthropic's words, "subject to change at Anthropic's discretion." Check the live page before you commit to anything annual.

Free gets you chat on web, desktop, and mobile. You also get web search and memory across conversations. Artifacts comes with it too.

What you don't get on Free: Claude Code. No Claude Cowork either, and none of Claude's strongest models. Free is a test drive. It was never built to be a daily driver.

Everything past Free splits into two different pricing logics, not five separate choices. Pro, Max, and Team charge a flat number you can put in a budget spreadsheet and forget about.

Enterprise and the raw API work the other way. They charge a number that moves with you, token by token, every single day you're logged in.

Most of the confusion you'll find elsewhere comes from people applying Enterprise logic to a Pro decision, or the reverse. Once you know which logic your plan runs on, the rest of this gets simple fast.

Here's the quick version before the detail. A single person who just wants Claude for daily work belongs on Pro or Max. A small group that needs shared billing belongs on Team. Anyone with a compliance checklist, or anyone building a product on top of the model, belongs on Enterprise or the raw API instead.

Pro and Max: the two plans built for one person

Pro costs $20 a month billed monthly. It costs $17 a month if you pay $200 up front for the year.

That's not a rounding trick. Annual Pro saves you $40 a year over paying month to month. Pay annually the day you're confident you'll stick with Claude past month one.

For that price you get Claude Code in your terminal. You get Claude Design, Slides, and Docs. You get Claude Science and more usage than Free. Opus access comes with it too.

Claude for Chrome and Claude for Microsoft 365 round out the plan. That matters if you want Claude reading your inbox or your browser tabs without a constant copy-paste routine.

The ceiling on Pro isn't a message count. It's a rolling five-hour session window. Anthropic has never published an exact number of messages, because the real constraint is token volume, and token volume depends on your model choice and how long a single conversation runs.

A short question to Haiku barely registers against that window. A long agentic coding session with Opus burns through the same five hours fast. Same plan, same price, wildly different runway.

That gap is what sends people to Max. If you hit Pro's session limit more than once or twice a week, you're not using Claude wrong.

You're using it at a rate Pro was never built to absorb.

Max fixes that with more usage, not new features. Max 5x costs $100 a month for five times Pro's per-session allowance. Max 20x costs $200 a month for twenty times Pro's allowance. Anthropic confirms both numbers directly: same models, same tools, a bigger bucket that refills every five hours and resets again each week.

Run the math before you upgrade. $200 a month works out to roughly $6.67 a day. If you're running Opus on multi-hour coding sessions and hitting Pro's wall before lunch most days, that's cheaper than paying the same token volume through the raw API.

If you hit Pro's limit once a week because of one unusually long session, you don't have a Max problem. You have a one-day problem.

Waiting five hours for the reset costs you nothing but fifteen minutes of patience. Paying $80 to $180 a month to avoid that wait is paying for a problem you don't actually have.

Team and Enterprise: where "per seat" stops meaning one price

Team is built for small groups, and Anthropic just lowered the bar for joining one. The current pricing page puts Team at "2 to 150" people. Earlier in 2026 the floor was five seats. A two-person shop now qualifies for the same admin controls that used to require padding the headcount just to hit the minimum.

Team runs two seat types, and you can mix them freely inside one org.

A Standard seat costs $20 a month per seat billed annually, or $25 billed monthly. It gives you more usage than Pro, plus single sign-on and domain capture. Central billing comes bundled with both.

A Premium seat costs $100 a month per seat billed annually, or $125 billed monthly, for five times a Standard seat's usage.

Both seat types currently include Claude Code, Claude Cowork, Claude Science, and Microsoft 365 integration, according to the live comparison table. Screenshot that table the day you sign your contract. Seat features on Anthropic's plans have shifted more than once this year, and a screenshot is the only record you'll have if they shift again on yours.

Enterprise is where the model stops resembling a subscription and starts resembling a utility bill. The headline number is $20 a seat a month, billed annually. That's deliberate. It's built to look identical to a Team Standard seat.

The part that isn't identical is usage. Team Standard usage is capped and included in your seat price. Enterprise usage is uncapped, billed separately, at the same per-token rates as the raw API.

Anthropic says this plainly in its own FAQ: Enterprise pricing is "$20 per seat per month plus usage billed at API rates, so cost scales with the models and tasks your team runs."

That sentence is the whole story. Walk through what it does to a real team.

Say you put ten people on Enterprise. The seat fee alone is $200 a month: ten seats at $20 each.

Seven of those people use Claude lightly, for drafting and quick research. Call it $10 a month each in token usage. That's $70.

Three of them run Opus against a large codebase most of the day. Call it $150 a month each in token usage. That's $450.

Add it up. $200 in seats, plus $70 in light usage, plus $450 in heavy usage, equals $720 a month for a team you budgeted at $200.

Same headcount. Same contract. A $520 gap the seat price never warned you about.

If your team's usage is steady and modest, Team Premium at a flat $100 to $125 a seat is easier to forecast than Enterprise's seat-plus-metered-usage math, even though Enterprise's headline number looks lower on the page.

Before you sign an Enterprise contract, pull a two-week usage sample from your three heaviest users on their current plan. Multiply that weekly token spend by 4.3 to estimate a monthly number, and build your budget from that figure, not from the $20 headline.

Read Enterprise's real cost from a usage report, not from the pricing page. The pricing page only tells you the floor.

What the Claude API actually costs, model by model

Underneath every subscription sits the same meter: tokens in, tokens out, priced per million. Anthropic's platform pricing page lists four current self-serve models this week.

Three cost levers sit on top of those base rates, and they're worth using from your first API call.

Prompt caching is the biggest one. Cache a repeated block of text, like a long system prompt or a codebase you keep referencing, and every repeat read costs a fraction of the base rate instead of the full price.

Run the math on Sonnet 5. Cache a 10,000-token prompt and reuse it across 100 requests. Paying full input price every time costs 1,000,000 tokens at $2 per million, or $2.00 flat.

With caching, the first request pays the write rate on those 10,000 tokens: $2.50 per million, so $0.025. The other 99 requests pay the cache-read rate of $0.20 per million on that same block: 990,000 tokens, or $0.198.

Total for the repeated context across all 100 requests: about $0.22. That's roughly 89 percent less for the part of your prompt that never changes.

Turn caching on anywhere your input has a stable prefix. A system prompt, a style guide, a reference document you quote in every call. The savings compound with every repeat.

The Batch API is the second lever: a flat 50 percent off both input and output for anything that doesn't need an answer in real time. Bulk summarization qualifies. So does overnight data extraction. The tradeoff is latency, not quality. Results land within 24 hours instead of instantly.

US-only inference is the third, required by some compliance setups. It adds a flat surcharge on top of the base rate. It doesn't replace the rate, so budget for both.

None of this touches your Pro, Max, or Team subscription bill. A Claude.ai subscription and a Claude API key are two separate accounts, billed separately, full stop.

Here's where that separation catches people. If you set an ANTHROPIC_API_KEY environment variable while running Claude Code on a paid plan, Claude Code can authenticate through that API key instead of your subscription. You end up paying per token on top of a subscription that was supposed to cover the work.

Anthropic's own Claude Code documentation gives you the exact way to check which one is actually charging you.

1. Open Claude Code and type /usage.
2. Read the "Total cost" line it prints, along with the per-model breakdown underneath it.
3. If that line shows a dollar cost against a model you expected your subscription to cover, type /usage-credits to see whether usage credits are already turned on.
4. On a Pro or Max plan, that command opens Settings > Usage on claude.ai in your browser, where you can switch usage credits off if you'd rather wait out a five-hour reset than pay automatic overage.

Run that check once a week if you're on a paid plan and also hold an API key for anything else. It takes thirty seconds. It's the only way to know which bill a given session actually landed on.

The limit nobody reads until they hit it

Every paid plan runs on the same clock. A five-hour session window resets on its own. A weekly cap sits on top of it, resetting once a week at a fixed time tied to your account.

Anthropic doesn't publish a fixed message count for any tier. The real constraint is token volume, and token volume depends on how long your conversations run and which model is doing the work.

That's not evasive marketing. It's the honest answer to a question that doesn't have a single number.

One question to Haiku and a ninety-minute agentic refactor with Opus both count as "a session" in name only. They consume the window at completely different rates. Two people on identical Pro plans can have two completely different experiences of the same limit, purely based on what they ask Claude to do.

Here's what actually happens when you hit the wall. You can wait for the reset, five hours out, and keep working the moment it lands.

You can upgrade to a higher tier instead. That takes effect immediately, and the price difference prorates against your current billing period, so you're never charged twice for the same day.

Or, on any paid plan, you can turn on usage credits. That lets you keep working past your included allowance, billed as overage at standard API rates.

That last option is the one worth understanding before you ever reach for it. Flip it on, and a flat-fee plan becomes a metered one for every hour you're over your cap. It's the same shift that makes Enterprise unpredictable, scoped down to the hours you're over instead of your whole account.

Leave usage credits off by default if budget certainty matters more to you than never waiting out a reset. Turn them on for a specific deadline week if you need to, then turn them back off the day after.

Set a spend cap the first time you enable them. Anthropic's Settings > Usage page lets you set a monthly limit on usage credits before you ever need to use that feature in a panic at 11pm the night before something is due.

Team plans extend the same overage mechanic to seats. A member who regularly burns through a Standard seat's allowance can run on usage credits instead of an automatic bump to Premium. That's the right call when the overage is occasional.

It's the wrong call when it happens every week. At that point the seat type itself is wrong, not the usage, and a Premium seat will cost less than the accumulated overage.

Which plan to actually buy

Match the plan to how you use Claude today. Not to what sounds impressive in a sales deck, and not to how you hope to use it in six months.

Evaluating Claude for the first time? Start on Free. Don't pay before you know if the model fits how you think.

Using Claude daily for writing, research, or light coding, and rarely feeling rushed? Pro, paid annually at $17 a month. The annual discount pays for itself in five months flat.

Hitting Pro's session limit once or twice a week, mostly from one long task? Stay on Pro. Wait out the reset. Upgrading to cover an occasional overflow means paying $80 to $180 a month extra to avoid a wait you'd forget about by dinner.

Hitting the limit most days, running Opus on real work? Max 5x at $100. Run the daily-cost math from the Pro and Max section above before you jump straight to Max 20x.

Coding against Claude for most of your working day, every day? Max 20x at $200, or a Team Premium seat if a teammate shares the load with you.

Running a 2-to-150-person group that needs shared billing and predictable usage? Team. Standard seats for the light users, Premium for the heavy ones, mixed in one org without penalty.

A large org with compliance requirements like SCIM, audit logs, or custom data retention, and usage that varies a lot by employee? Enterprise, with a usage budget pulled from a real sample and watched from month one.

Don't discover your real number in month three, the way the ten-person example above did.

Building a product that calls Claude programmatically? Skip every subscription tier entirely. Go straight to the API, priced by the token, with caching and batching turned on before you send your first real request.

The most expensive mistake isn't underpaying on Free. It's jumping straight to Max or Enterprise because the name sounds like what a serious user should have, before measuring how often the cheaper tier's limit actually gets in your way.

Track your own usage for one week on whatever plan you're already on. Open Settings > Usage, check how close you came to your limit each day, and let that number pick your next plan instead of a feeling.

What Claude pricing doesn't tell you: the real bill is bigger than one model

Here's the part Anthropic's pricing page was never going to mention, because it isn't Anthropic's problem to solve. Almost nobody uses only Claude.

The same week you're deciding between Pro and Max, you're probably also paying for ChatGPT, because Claude occasionally stumbles on something OpenAI handles better.

Maybe Gemini too, for its longer context window. Maybe Grok, because somebody on your team swears by it for one specific task.

Subscribing separately to ChatGPT Plus, Claude Pro, Gemini Advanced, SuperGrok, and Perplexity Pro runs about $100 a month. None of those five tools remembers a single thing you told the other four.

That's the real cost of Claude pricing in isolation. It's never just the $17 or the $20 or the $100. It's that number, plus three or four more identical numbers for the other models you keep paying for separately, plus the hidden tax of re-explaining your project every time you switch tabs.

What an AI aggregator actually is comes down to one test: does your context survive when you switch models mid-conversation, or does every new tab start you from zero.

How to switch between AI models in one workflow without losing that context is the entire reason bundling beats stacking logins.

Magai runs Claude alongside ChatGPT, Gemini, Grok, DeepSeek, and dozens more models inside one workspace, with your thread's context intact when you switch. Standard runs $20 a month, the same price as Claude Pro alone, and it includes every standard model instead of one.

Still deciding whether Claude or ChatGPT deserves your money first? Claude vs ChatGPT: which one should you pay for walks through where each one wins.

Want the direct tradeoff against Magai itself? The full Claude vs Magai comparison breaks it down line by line.

For the rest of how the frontier models compare against each other as pricing shifts every few months, read the posts filed under Artificial Intelligence. That category tracks every major model's pricing and capability changes as they happen, not just Claude's.

Stacking five separate AI subscriptions instead of bundling them is the single most common way people overpay for AI in 2026. It has nothing to do with which pricing page you read most carefully.

Claude's pricing, on its own, is honest once you read past the headline number. The $20 version of Claude really is $20.

The mistake isn't misreading Anthropic's page. The mistake is treating Claude's price as the whole bill.

For most people it's just the first of several identical charges, stacking up quietly on a card statement nobody adds up until the renewal emails all land the same week.

Pick the Claude tier that matches how you use it this week. Then open your card statement and add up every other AI subscription sitting next to it, before you renew a single one of them on autopilot.

If the total surprises you, see what one workspace instead of five actually costs.