Cursor request limits, explained
Cursor's plans combine an included allowance with usage-based pricing once it runs out, so 'hitting the limit' rarely means being locked out — it means starting to spend money. This page covers where the requests actually go and how to keep them down.
Short answer
The number that matters is not how many times you typed something — it is how many model calls your instruction caused. Cursor's agent works in steps: read a file, propose an edit, run a check, read the result, try again. Each of those steps is billable work. A single 'refactor this module' can therefore spend a dozen or more requests, which is why a monthly allowance that sounds generous can evaporate in an afternoon of agent-heavy work.
The cap on each plan
Read the window column before the cap column. A cap that resets on a rolling window behaves nothing like one that resets on the 1st of the month, and it is the reason two people on the same plan report completely different experiences.
Hobby / Free
$0Cap
A small monthly allowance of agent requests plus a limited number of completions, sized for evaluation rather than daily work.
Window
Calendar month
When you hit it
Agent features stop until the month turns over. Basic editing continues — you lose the assistant, not the editor.
Pro
$20/moCap
An included monthly usage allowance, after which you can either stop or opt into usage-based pricing and keep going at cost.
Window
Calendar month, aligned to your billing date
When you hit it
Nothing breaks silently. You are prompted to enable usage-based pricing; decline and agent features pause until the next cycle.
Ultra
$200/moCap
Published as a large multiple of the Pro allowance, aimed at people whose agent runs all day.
Window
Calendar month
When you hit it
Same overflow mechanic as Pro. The tier exists to make heavy agentic use predictable rather than unlimited.
Published figures read off Cursor's own documentation in August 2026. Caps on this category of product change without announcement and are often adjusted per-account under load, so treat the mechanics below as the durable part and Cursor's own pricing and usage documentation as authoritative over anything on this page.
What actually consumes one unit
This is the part that lets you predict your own cap instead of being surprised by it.
One agent instruction is many requests
The agent reads files, plans, edits, runs checks and reacts to output. Each of those round-trips is a model call. This is the single biggest gap between what people expect to spend and what they actually spend.
Model choice changes the cost per request enormously
Frontier reasoning models cost multiples of the cheaper defaults for identical work. Leaving the most expensive model selected for boilerplate is the most common way to burn an allowance without noticing.
Context size is the multiplier on everything
Every file the agent pulls into context is re-read on subsequent steps. Pointing it at a directory rather than a file can multiply the cost of the same task several times over.
Tab completions are metered separately from agent requests
Autocomplete draws on its own allowance, so exhausting agent requests does not necessarily stop completions, and vice versa. Check which one you actually ran out of before upgrading.
Failed and retried steps still count
An agent run that goes down a wrong path, notices, and backtracks has spent every request along the way. Long autonomous runs on ambiguous instructions are where allowances quietly disappear.
How the reset really works
- •The allowance is monthly and tied to your billing date rather than the 1st — check the billing page before planning a heavy week.
- •Usage-based pricing is opt-in. If you have not enabled it, you cannot be surprised by a bill; you will simply stop having agent access until the cycle turns.
- •If you have enabled it, set a spending cap the same day. It is the only thing standing between a runaway agent loop and a genuinely unpleasant invoice.
- •The in-product usage dashboard is authoritative and updates as you work. Any figure published elsewhere, including this page, is a snapshot.
- •Cheaper or auto-selected models often draw from a separate, much more forgiving pool than the frontier models do — worth checking before assuming you are out of quota entirely.
Getting more out of the plan you already pay for
Try these before upgrading. For most people one or two of them push the cap far enough out that the next tier stops being necessary.
Scope the agent to files, not folders
Naming the two or three files that matter, instead of letting the agent search a directory, is the difference between a three-request task and a thirty-request one. Nothing else on this list saves as much.
Drop to a cheaper model for mechanical work
Renaming, boilerplate, test scaffolding and formatting do not need frontier reasoning. Reserve the expensive model for genuine architecture and hard debugging.
Write the instruction as a spec, not a wish
Ambiguous instructions cause exploratory runs, and exploration is what costs. Stating the file, the intended behaviour and the constraint up front routinely halves the number of steps a task takes.
Stop a run that is going the wrong way
Watching an agent thrash for another five steps because it might recover is pure spend. Interrupt it, correct the instruction, and restart — the restart is cheaper than the recovery almost every time.
Use inline edits for small, known changes
If you already know exactly what to change, a targeted inline edit costs one call. Handing the same change to the agent costs a planning step, an edit step and a verification step.
If most of your spend is reasoning about code rather than editing it, an assistant subscription covers that far more cheaply than agent runs do.
Upgrade, or move?
Pay for the higher tier when…
- ✓You run the agent most of the working day — the higher tier makes heavy use predictable, which usage-based overflow does not.
- ✓Your overflow spend has quietly exceeded the price difference between tiers for two months running. Check the usage dashboard before assuming it has not.
- ✓You need frontier models specifically for architectural work, so downgrading the model is not an available saving.
Look elsewhere when…
- →You mostly want inline completions rather than an autonomous agent — that is a much cheaper category of product and you are paying for capability you never invoke.
- →Your organisation already pays for a coding assistant bundled elsewhere; running two overlapping subscriptions is the most common avoidable cost in this category.
- →The agent's value to you is really conversational — reasoning about design rather than editing files — in which case a general assistant subscription covers it for less.
Where the quota is looser
- GitHub Copilot — Strongest on inline completion and frequently already bundled with an existing developer plan. Cheaper if you rarely invoke an autonomous agent.
- Claude — If most of your usage is reasoning about code rather than editing files in place, an assistant subscription covers it at a fraction of agent-run cost. Its limits page is linked below.
- ChatGPT — A reasonable overflow destination for design discussion and debugging chat while your coding allowance recovers.
Common questions
What counts as one request in Cursor?
One model call — not one thing you typed. The agent reads files, plans, edits, runs checks and reacts to the results, and each of those steps is its own call. A single instruction like 'add tests for this module' commonly spends between five and twenty requests depending on how much context it has to pull in.
What happens when I run out of requests on Pro?
Nothing breaks silently. You are offered usage-based pricing, which lets you keep working and pay for the overflow. If you decline, agent features pause until your billing cycle turns over while ordinary editing continues. If you accept, set a spending cap immediately — it is the only guard against a runaway agent loop.
Why did my usage spike without me doing anything differently?
Usually one of two things: a more expensive model was selected, or the agent pulled a much larger context in — a folder instead of a file, or a long session that accumulated open files. Both multiply the cost of identical work. The usage dashboard breaks spend down by model, which normally makes the cause obvious within a minute.
Do tab completions use the same allowance as agent requests?
No, they are metered separately. It is common to exhaust one while the other still has capacity, so check which one actually ran out before concluding you need a bigger plan.
Is Ultra worth it, or should I just enable usage-based pricing?
Compare two months of actual overflow spend against the tier difference — the usage dashboard has the number. Overflow is better when your load is bursty; the higher tier is better when the agent runs all day, mostly because it makes the cost predictable rather than because it is cheaper per request.
Related reading
People hitting Cursor's limits ask AI what to move to. Is that your tool?
"I keep hitting the Cursor cap, what should I use instead?" is one of the highest-intent questions asked of an assistant — the reader is already paying and already frustrated. ChatGPT, Claude and Perplexity answer it from pages like this one. If you build a tool with a more generous quota, check whether they name you.
📬 Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.