Cursor AI pricing: what the plans cost and what I paid

Contents

How much does Cursor cost?

Cursor’s sticker price is $20 a month. Across eleven months I paid $510.70.

Nothing went wrong and nobody overcharged me. The gap between those two numbers is the whole subject of this page, because Cursor’s bill has two halves and only one of them appears on the pricing page.

C

Here is the full ladder, read off Cursor’s own pricing page in a US browser on 11 August 2026, with the billing toggle switched both ways and every nested sub-tab opened.

Plan	Monthly	Yearly (per month)	Agent limits
Hobby	Free	Free	Limited requests, no card
Pro	$20	$16	The baseline
Pro+	$60	$48	3× Pro
Ultra	$200	$160	20× Pro
Teams — Standard	$40 per user	$32 per user	Pro-equivalent, pooled
Teams — Premium	$120 per user	$96 per user	5× Standard
Enterprise	Custom	Custom	Quote-only

Yearly billing takes exactly 20% off every paid tier. Not “about 20%”, and not more on the expensive plans — the same fifth, all the way up.

Why published Cursor prices disagree with each other

If you have already searched this, you will have noticed the summaries do not match. One says six tiers from free to $200. Another says the seats are $40 and $120. A third lists Pro, Pro+ and Ultra and never mentions seats at all.

They are all quoting the same page. The reason they conflict is that Cursor’s pricing page is nested: four cards across the top, and two of those cards hide sub-tiers behind tabs you have to click. The Individual card shows $20 until you click Pro+ or Ultra, at which point the same card shows $60 or $200. The Teams card shows $40 until you click Premium and it becomes $120.

A tool that fetches the page gets whichever tab happened to be selected. That is how a page ends up confidently reporting that Cursor Teams costs $40 per user, full stop, when there is a $120 seat one click away.

It is worth knowing this beyond the trivia, because it tells you something about how to check any price on this page: the number you see first is the cheapest member of a set, not the price of the thing.

What the bill is actually made of

Every paid plan is two charges wearing one price tag.

The subscription is the number in the table above. It buys you the editor, unlimited Tab completions, and an allowance of frontier-model usage denominated in dollars.

The usage is what happens after the allowance runs out. On an individual plan Cursor bills the model’s API rate with no markup of its own, and it bills in arrears, after the month in which you spent it. On Teams and Enterprise there is a markup, and I had this wrong until I read the docs: third-party models carry an additional Cursor Token Rate of $0.25 per million tokens, and choosing regional data residency adds a further 10% uplift on model pricing.

That second word is where the money hides. Usage-based billing is not unusual and Cursor is unusually clean about it: no marked-up tokens, no invented credit currency with a fictional exchange rate, a published rate card. What makes it bite is the delay. You do not feel a meter running. You feel an invoice, four weeks after the work that caused it, by which point you cannot connect the number to the afternoon that produced it.

There is a lever for this and it is the most important sentence on this page: Cursor offers a spend limit on Pro, Pro+ and Ultra. It caps what usage-based billing can add. It is off unless you turn it on.

What I actually paid, from the invoices

I paid for Cursor for eleven months and then cancelled. Here is the whole of it, from the Stripe invoice history rather than from memory.

	Amount
Subscriptions across eleven months	$340.00
Metered usage on top	$170.70
Total	$510.70
Worst single month	$159.22
Usage as a share of the total	33%

I joined at $20 a month. In October I upgraded to Pro+ at $60, specifically to stop the overages — which is worth sitting with, because it did not work. Buying a bigger allowance moved where the line was without changing what happened when I crossed it.

A third of what I paid Cursor never appeared on any pricing page I read before signing up.

The month that made the point was October 2025: $159.22, on a plan advertised at $20, with roughly half of it unplanned. Nothing was broken. I had simply had a heavy month and the bill arrived a month later.

Is this Cursor’s fault?

Mostly not, and it matters to be precise about it, because “Cursor is expensive” is the wrong lesson.

Cursor charges the model’s API rate with no markup on individual plans. It publishes the rates. It offers a spend limit on every individual paid tier. Three of those are more than several competitors manage, and the fourth — the spend limit — would have capped my entire problem at whatever number I chose.

I never set it. That is not a defect in the product; it is a default I did not change and a bill I did not read closely enough for several months.

What is fair to put on Cursor is the shape: a headline price that is a floor rather than a price, and arrears billing that puts the feedback a month behind the behaviour. If the meter showed a running monthly total the way the plan page shows $20, most of this page would not need writing.

So the honest framing is not that Cursor overcharges. It is that Cursor’s advertised price is not a prediction of your bill, and the tool gives you exactly one mechanism to make it one.

How Cursor’s usage actually works

The word “credits” does a lot of hiding, so here is the machinery underneath it, from Cursor’s own models-and-pricing documentation.

Your allowance is two separate pools that both reset monthly, not one balance:

• Cursor’s own models (Grok 4.5 and Composer 2.5) draw on their own pool.
• Everything else (Claude, GPT, Gemini) draws on a third-party pool charged at the model’s API price.

That split matters because the two run down independently. Exhausting one does not touch the other, and “I’m out of credits” usually means you are out of one specific pool.

	Cursor’s own models	Third-party models
Covers	Grok 4.5, Composer 2.5	Claude, GPT, Gemini
Charged at	Its own pool	The model’s API price
Resets	Monthly	Monthly
Markup on individual plans	None	None
Markup on Teams / Enterprise	None	+$0.25 per million tokens

What each tier includes is the number nobody quotes. Paid individual plans bundle between $20 and $400 a month of third-party model usage, scaling up the ladder. Read that against the subscription prices and the top of the range gets interesting: Ultra costs $200 a month and includes $400 of third-party usage. Whether that is a saving depends entirely on whether you would otherwise have spent it.

What is free. Tab completions are unlimited and consume no usage at all, on every paid tier. If most of your day is autocomplete rather than agent runs, your bill will look nothing like mine.

Auto mode versus picking a model. Auto Cost is priced per million tokens regardless of which model it routes to — you are buying a rate, not a model. Auto Balance and Auto Intelligence charge the actual API rate of whatever they choose. Selecting a frontier model by hand charges that model’s standard rate. This is the single biggest lever on the bill, and it is one dropdown.

Extended context is a multiplier, not a feature. GPT-5.4 and GPT-5.5 charge 2× input pricing for long-context requests up to a million tokens. On legacy request-based plans, Max Mode adds 20% to the model’s API rate in exchange for a larger context window. Neither is a trap — both are documented — but “just turn on the big context window” is a pricing decision.

What the models themselves cost

Since the third-party pool is billed at API rates, the model you pick is the price you pay. A sample from Cursor’s published rate card:

Model	Input / M tokens	Cache read / M	Output / M
Claude Sonnet 5	$2.00	—	$10.00
GPT-5.4	$2.50	$0.25	$15.00
Gemini 3.6 Flash	$1.50	$0.15	$7.50

Two notes before you plan around these. The Claude Sonnet 5 rate is promotional through 31 August 2026, so it is a discount with an expiry rather than a price. And cache reads are an order of magnitude cheaper than fresh input, which is why a long conversation in one thread costs less than the same work restarted repeatedly — the cache is doing the work.

The spread between Gemini 3.6 Flash and GPT-5.4 is roughly 1.7× on input and 2× on output. Across a heavy month that is not a rounding difference, and it is entirely under your control.

Five ways to spend less on Cursor

None of these are clever. They are the things I did not do.

Tactic	Where it lives	What it changes
Set a spend limit	Account billing settings	Caps metered usage outright. Off by default
Stay in Auto for routine work	Model dropdown	Flat per-million rate instead of frontier prices
Match model to task	Model dropdown	Up to ~2x on output between Gemini Flash and GPT-5.4
Treat long context as a purchase	Per request	Avoids 2x input pricing you did not need
Check consumption mid-cycle	Usage dashboard	Turns a month-late surprise into a live number

Set the spend limit. It exists on Pro, Pro+ and Ultra, it caps what usage-based billing can add, and it is off until you turn it on. If you do nothing else on this list, do this one. My entire $170.70 of overage would have been whatever number I typed into that box.

Stay in Auto for routine work. Auto Cost buys a flat per-million-token rate regardless of which model answers. For refactors, boilerplate and the fiftieth small edit of the day, the frontier model is not earning its premium.

Match the model to the task, not to your mood. Reaching for the most expensive model by default is how the bill grows without the work improving. The rate card above is the whole argument.

Treat extended context as a purchase. 2× input pricing for long context is a fair trade when you genuinely need a million tokens of it, and pure waste when the answer was three files away.

Check consumption mid-cycle. Arrears billing means the invoice is the last place you learn what you spent. The dashboard knows in real time; the bill knows a month late. Look at the one that is current.

One line item that is easy to miss: Bugbot. On individual plans the pricing page lists it as “Bugbot on usage-based billing” rather than as an included feature, which means Cursor’s automated code review draws on the meter like anything else. On Teams it appears as “agentic code reviews with Bugbot”, bundled into the seat. So the same capability is metered on a personal plan and included on a team one, and if you turn it loose on a large pull request from a Pro account, that review has a price.

What Cursor costs a team

Team pricing is where the nested tiers matter most, because the gap between the two seats is 3×.

Standard at $40 per user is Pro-equivalent access plus the administrative layer: centralised billing, a team marketplace for rules and plugins, usage analytics, team-wide privacy mode, and SAML/OIDC SSO.

Premium at $120 per user adds 5× the Standard Agent limits, agentic code reviews with Bugbot, and cloud agents with shared team context.

For ten developers that is $4,800 a year against $14,400 at list, or $3,840 against $11,520 billed yearly — before any usage. Usage pools across the team on Enterprise, which changes the arithmetic again and is one of the reasons Enterprise is quote-only. Team billing also carries the markup individual plans do not: $0.25 per million tokens on third-party models, on top of the model’s own rate, and 10% more again if you need regional data residency.

The trap at team scale is the same one I hit personally, multiplied. A per-seat subscription is predictable and a metered overage is not, and ten people crossing their allowance in the same busy month produces a bill nobody forecast. If you are buying seats, the spend controls are not an optimisation to get to later. They are the thing that makes the budget a budget.

How Cursor pricing compares

Against the tools I have tested on identical work, Cursor sits in the middle on price and at the top on variance.

Tool	Entry price	Free tier	Billing shape
Cursor	$20/mo	Yes, evaluation-grade	Subscription plus metered usage
Claude Code	$20/mo	No	Subscription, or metered API
Google Antigravity	Free	Yes, does real work	Free, or a Google One plan
Codex	Bundled with ChatGPT	Yes	Included in the subscription
Lovable	$25/mo	5 credits a day	Credits, which expire

Two things stand out. Cursor is the only one of these where the entry price and the likely bill are different numbers by design. And Antigravity is the only one whose free tier will carry real work indefinitely — I built a complete app on it without paying anything, which makes it the cheapest way to find out whether you like working this way at all.

That is not an argument that Cursor is bad value. Eleven months of it were genuinely useful and I would have kept paying if the bill had been predictable. It is an argument that the comparison people make — “$20 versus $20” — is not the comparison that shows up on the card.

One thing to check before you commit to a year

Cursor’s parent company, Anysphere, is being acquired by SpaceX in a $60 billion all-stock deal announced on 16 June 2026 and expected to close in the third quarter of 2026 — which is to say, around now. It is the largest startup acquisition on record, and it followed Cursor reaching roughly $2 billion in annual recurring revenue.

I am not going to speculate about what a new owner does to a price list. What I will say is narrower and checkable: this pricing structure has already been rewritten once, when Cursor moved from a request-counting model to dollar-denominated credits in mid-2025, and that change was contentious enough to still be the subject of explainer posts a year later.

So the yearly discount is a real 20%, and it is also a twelve-month commitment to a plan structure that has changed before, under an owner that is new. If your usage is heavy and stable, take the discount. If you are still working out whether Cursor fits, paying monthly for a few months costs you a fifth more and buys you the ability to leave. I would take that trade today.

Which Cursor plan should you pick?

Hobby, free — if you are evaluating. It asks for no card and it will tell you within a week whether the editor suits you. It will not sustain daily work, and it is not meant to.

Pro, $20 — the default, and the right starting point for one developer. Do not treat it as your monthly cost; treat it as your floor, and set a spend limit the day you subscribe.

Pro+, $60 — worth it only if you have watched yourself exceed Pro’s allowance repeatedly. Buying a bigger allowance to escape overages is the mistake I made; it raises the ceiling without changing the mechanism. Upgrade because 3× the limits matches your measured usage, not as a way to stop being surprised.

Ultra, $200 — for all-day agent use where 20× Pro’s limits is a genuine requirement. At this price the question is no longer whether Cursor is worth it but whether a flat $200 beats metered usage at your volume, and that is answerable only from your own invoices.

Teams, $40 or $120 per user — Standard for the admin and privacy layer, Premium if your team lives in the agent. Decide on the 5× limits, not the feature list.

Students — there is no standing student plan, whatever you may have read. Cursor’s students page offers the same free tier as everyone else, plus event promotions with no published terms. Budget as though the discount does not exist, because today it does not.

What the free Hobby tier actually gets you

“Is Cursor free” is one of the most common things people search alongside the price, and the answer deserves more than yes.

There is a free tier, it is called Hobby, and it does not ask for a card. What Cursor’s pricing page lists against it is short:

Hobby includes	Hobby does not include
No credit card required	Extended Agent limits
Limited Agent requests	Generous limits for Grok and Composer
Access to Composer	Access to frontier models
	MCPs, skills and hooks
	Cloud agents

The word doing the work there is limited. Cursor does not publish what the Hobby Agent allowance actually is, in requests or in dollars, which puts it in the same position as several rivals: a free tier you can only size by hitting the end of it.

What I can tell you is one measured data point. The Stack Cost build above — six requirements, 581 lines, done in under a minute — ran entirely on Hobby and cost nothing. So the free tier is not decorative. It will complete a real, self-contained task.

What it will not do is the thing that generated my $510.70: all-day iterative work against frontier models, which is precisely what the paid tiers open up and precisely what the Hobby feature list excludes. Note that “access to frontier models” is a paid line item. On Hobby you are working with Composer and Cursor’s own models, not Claude or GPT.

There is no separate free trial of the paid tiers, either. The free tier is the trial, and the upgrade is a purchase rather than a countdown.

Does Cursor have a student discount?

This one is worth its own answer, because the web is fairly confident about a thing that does not appear to be true.

Several guides list a Cursor student discount as an established offer, sometimes with terms attached. Cursor’s own students page does not describe one. It says anyone can get started with Cursor for free, and directs students to watch for promotions at on-campus and online events, with a campus newsletter for early notice.

That is not a discount programme. It is a free tier everyone already has, plus the chance of a time-limited promotion with no published terms, no stated eligibility, and no verification flow.

If you are a student budgeting for this, budget for Pro at $20 or the free Hobby tier, and treat any promotion you catch as a bonus rather than a plan.

What actually happened to Cursor’s pricing in 2025

A good share of the people searching for Cursor’s price are really asking what happened to it, and the answer explains the shape of the bill you will get today.

Cursor originally sold requests: a monthly quota of a countable thing, where 500 fast premium requests meant 500 of them and you could feel the number going down. In mid-2025 it moved to dollar-denominated usage — your plan includes an amount of model spend rather than a count of actions.

	Before mid-2025	After
You bought	A count of requests	An amount of model spend
A unit was	1 request, always	$1 of usage, size varies
Predictable?	Yes, countable	Only if you watch it
Overspend showed up	When the quota ran out	On next month’s invoice

The change was contentious enough that explainer posts about it are still being written a year later, and the reason is not really the money. A request is a unit you can hold in your head. A dollar of frontier-model usage is a unit whose size depends on which model answered, how long the context was, and whether the cache was warm. The same action can cost different amounts on different days, and no amount of documentation makes that as legible as “you have 500 of these.”

It is worth saying that the new model is more honest than the old one. Charging for what you actually consume, at published rates, with no markup on individual plans, is a better-behaved arrangement than a quota whose real cost is hidden inside an average. But it moved the mental effort onto the customer, and a pricing model you have to think about is one people will describe as having got worse even when it got fairer.

That history is why the advice on this page keeps coming back to the same two habits: set the cap, and look at consumption before the invoice does.

What a real build cost me on Cursor

Abstract rates are hard to weigh, so here is one concrete data point from the same brief I have given seven AI coding tools.

I asked Cursor to build a single-page subscription tracker with six specific requirements — add a tool with a name, cost and billing cycle; show annual plans as a monthly equivalent; a running total pinned to the top; a cancelled state that leaves the total but stays visible; localStorage persistence; a dark theme.

It met all six in under a minute, writing 581 lines across three files, on the free Hobby tier, for nothing. It was the fastest of the seven tools I have run that brief through, and the tightest apart from Codex’s 585 lines.

That is the case for Cursor’s pricing that the invoice story obscures. A single well-specified task, done in one pass, costs approximately nothing — the free tier absorbed it entirely. My $510.70 did not come from tasks like that. It came from eleven months of open-ended, iterative, all-day work, which is a different activity that happens to use the same editor.

Which tells you how to read your own likely bill. If your use looks like discrete, specifiable jobs, Cursor is cheap and the free tier may genuinely cover you. If it looks like living in the thing, assume the meter, and set the cap.

The final word

Cursor’s pricing is honest and its presentation is not.

Every number is published, the model rates carry no markup, the spend limit exists, and the yearly discount is a clean 20% with no tier games. That is a better-behaved price list than most of this category manages.

But the number that gets quoted is $20, and the number that arrives depends on a meter you cannot feel, reported a month late, with the one control that would fix it switched off by default. My eleven months came to $510.70 — two and a half times the subscription total I thought I was signing up for — and every dollar of that gap was my own usage, billed fairly, that I simply could not see accumulating.

If you take one thing from this page, take the boring one. Subscribe at $20, and set the spend limit before you write any code. It is two minutes, it is on every individual paid tier, and it is the difference between a price and a bill.