Cursor says "You've hit your usage limit": which pool fired

The exact wording, and which one you are looking at

Cursor does not have one usage-limit message. It has a small family of them, and which one you are reading already narrows the diagnosis before you open anything. These are the strings shipped in the current desktop build, checked against Cursor 3.15.19 on 19 August 2026.

cursor total usage limit reached

People searching cursor total usage limit reached are in the same family as "You've hit your usage limit". The banner still does not name the empty pool.

Cursor has two paid agent pools plus a Hobby ceiling. Cursor Models and Other Models drain independently. Hobby Tab and Hobby Agent are their own limited meters, not this paid banner. Identify which balance is empty, then wait, route to the other pool, or Get Plus when you need overflow.

What about "You've hit your free requests limit"?

This one deserves its own answer, because it is the message people quote most often and it is no longer in the product. "You've hit your free requests limit" belongs to Cursor's request-based era, when free and paid plans were counted in fast and slow premium requests rather than usage. Searching the current desktop build for it finds nothing: the literal phrase "requests limit" does not appear in Cursor 3.15.19 at all.

So if you are reading that sentence today, one of three things is true, and they need different responses.

Next, identify which limit you hit

Do this before clicking Upgrade. Keep the failed conversation open and record four facts: the exact wording, the selected model, the plan shown in the dashboard, and whether Tab still works. That small evidence bundle separates account capacity from a local editor fault and makes any support request useful.

What triggers the message

The paid third-party pool is the common professional case. Cursor values each request at the selected model's API rate. Agent runs with long context, large outputs, repeated tool calls, or several self-correction turns can consume far more than a short question. A user can therefore hit the limit after fewer visible prompts than last month even if total working hours look similar. The prompts were not economically equivalent.

The free Tab case is different. Cursor can request a prediction when typing pauses, and the quota is not merely a count of suggestions accepted with the Tab key. This explains the apparently impossible state where the free account shows few accepted completions but predictions stop. Paid plans make Tab unlimited, but that does not make paid Agent usage unlimited.

A team cap is different again. A seat may still have plan capacity while a per-user or organization policy denies spend. Only the admin view can reconcile that. Screenshots should include the cycle date and pool label, not only a large percentage or a red banner.

Cursor Pro usage limits, specifically

Pro is where most of these reports come from, so it is worth separating what Pro actually includes from what people assume it includes. At $20/mo it is the baseline paid tier, and the thing to understand is that it does not buy one allowance. It buys three different ones that empty independently.

The $20 in that third row is the number that surprises people, because it is the same figure as the plan price and it is not the plan price. It is a dollar-denominated allowance of third-party model usage, billed at what those vendors charge. A pinned frontier model on long context can spend a meaningful fraction of it in a single agent run, which is why two months with similar hours can end very differently. Pro Plus at $60 raises that allowance to $70 and Ultra at $200 raises it to $400.

One structural point that decides whether Pro is the right plan for you at all: Cursor allowances refresh monthly and unused usage does not roll over. A quiet month does not bank capacity for a heavy one. If your work is spiky, that argues for staying on Pro with a small deliberate on-demand cap rather than upgrading to a tier sized for your worst week.

When the limit resets on each plan

Cursor allowances refresh monthly on the account's billing cycle. They do not use a daily five-hour window, and the cycle is not assumed to start on the first of the calendar month. If a Pro subscription began on the 17th, the relevant date is normally the account renewal date around the 17th. Read the date in the dashboard or Manage Subscription rather than calculating from memory.

If the dashboard says the cycle has reset but the same model still reports exhaustion, capture the cycle dates, remaining balance, exact error, and time. Then fully quit and reopen once to refresh account state. If the mismatch persists, it is no longer normal quota behavior. Send the evidence to Cursor support. Do not repeatedly delete local state; a cached enforcement or account entitlement issue is server-side and destructive local troubleshooting adds risk without changing the account.

The immediate workarounds, in order

The best workaround preserves the task without silently converting it into an unlimited bill. Move down this list and stop as soon as one option matches the work.

1. Switch to an eligible model in the other pool. For routine implementation, Composer or another included Cursor model may be entirely adequate when the third-party allowance is empty.
2. Reduce the task before retrying. Start a fresh conversation, name the files, and ask for one bounded change. This will not refill a pool, but it avoids spending overage on irrelevant context.
3. Use Tab or Inline Edit for mechanical work. Paid Tab is unlimited and often better than Agent for changes you can point at precisely.
4. Enable capped on demand. Choose a small hard ceiling that buys the deadline rather than an open-ended continuation for the month.
5. Upgrade if this is a repeated normal month. Compare the actual last two cycles of overage with the signed-in price of the next tier.
6. Fail over to another tool you already fund. A Claude Code or Codex subscription has an independent quota. Keep it in a separate worktree if another agent may still be editing the repository.
7. Wait for reset. This is the correct answer when budget is fixed and the blocked work is not urgent.

Changing models mid-conversation deserves one caution: the new model inherits the existing conversation context. A very long chat can remain expensive even after the route changes. For a clean failover, summarize the task, include the current repository state and the exact remaining step, then start a fresh chat. Review any existing uncommitted diff before handing it to another agent.

What will not fix a usage limit

Quota is attached to the account. Local troubleshooting is useful for a slow extension host, a corrupt chat database, or broken network streaming. It does not create subscription capacity. The following rituals waste time specifically when the dashboard confirms an exhausted pool.

• Reinstalling Cursor. The new binary signs into the same account and reads the same server-side balance.
• Deleting caches or chat history. This can improve local performance, but it cannot reverse usage that has already been metered.
• Signing out and back in repeatedly. One refresh after a cycle boundary is reasonable; repeated authentication is not a quota mechanism.
• Changing networks or disabling a VPN. Good for stalled streams, irrelevant to a clear account-limit response.
• Retrying the same expensive prompt. If the error is ambiguous and requests partly succeed, retries can create more usage rather than less.
• Creating extra accounts. This evades rather than solves policy, fragments history, and may violate organizational expectations.
• Trusting an old request-count article. Fast and slow premium requests belong to a retired pricing system.

There is one edge case: the balance visibly refreshed, yet enforcement did not. In that case fully quitting once, reopening, and signing in once can refresh the client view. If the server still rejects the request, escalate with evidence. The distinction is temporal: before the displayed reset, local actions are irrelevant; after the displayed reset, one refresh is a diagnostic step.

How to choose between waiting, overage, and upgrading

Use costs already incurred, not hypothetical request equivalents. If the last two months needed $18 and $24 of on demand above Pro, compare that actual range with the current price and allowance of Pro Plus. If the overage arose from a one-time repository migration, do not annualize it. If it arose from ordinary daily Agent runs, do.

Also price interruption. An engineer waiting half a day can cost more than a modest overage, while a hobby project can wait at no meaningful cost. This is why a single recommendation such as "always enable on demand" or "never pay overage" is not serious. The meter is the same; the consequence of reaching it is not.

Monitor headroom so it stops being surprising

The number that prevents a limit incident is not spend to date. It is remaining headroom relative to time left. A dashboard showing $12 consumed is incomplete without the plan allowance, the cycle start, the reset date, and whether usage accelerated this week.

Cursor's own dashboard is authoritative for what Cursor will enforce. A cross-agent view answers a different question: whether the work can fail over to another subscription and what the combined workload is doing. Do not replace the vendor balance with an estimate when making a billing decision. Use the estimate for forecasting and the vendor dashboard for entitlement.

A five-minute incident checklist

Time and timezone:
Cursor version:
Plan and seat type:
Exact error text:
Selected model:
Feature that failed: Tab / Agent / Cloud Agent / other
Feature or model that still works:
Included usage shown:
On-demand usage and hard cap:
Billing-cycle reset date:
Personal or managed team account:
Cursor status at the time:
Request ID, if the failed conversation provides one:

That checklist does two jobs. It makes the immediate decision obvious, and it prevents support from spending the first exchange asking for information that was visible when the incident happened. A request ID is especially useful for an ambiguous provider or streaming failure. It is not necessary to prove a dashboard balance is empty.

Questions people ask

Keep reading

Continue through Errors and fixes with Why Cursor is slow, and how to isolate the real cause.