Cursor Pricing Limits Explained: How Far $20 Goes
Cursor pricing confuses everyone. Here is exactly how many requests the $20 Pro plan buys, what usage credits mean, and when overages start.
By Capital & Compute
Cursor’s paid plans no longer sell a fixed number of requests. Each tier is a monthly pool of model usage measured in dollars: $20 on Pro, $60 on Pro+, $200 on Ultra, all spent at the model makers’ API rates. So $20 buys roughly 225 Claude Sonnet requests, or about 650 on a cheaper model. The request count is not a plan feature anymore. It moves with the model picked and how heavy each task runs.
That single change is the source of nearly every “what happened to Cursor pricing” complaint. The old plan promised a round number of requests. The new one promises a round number of dollars, and dollars buy very different amounts of work depending on the model. This is what the tiers actually include, how far each pool goes, and where the surprise bills come from.
What actually changed
For most of its life Cursor sold a request count. The Pro plan was widely understood as 500 fast requests a month. That number was easy to reason about: one prompt, one request, 500 of them.
In its 2025 pricing clarification, Cursor (the product of Anysphere) moved Pro to a usage-credit model: a $20 monthly pool of model usage charged at the underlying API rates. The company’s stated reason is that newer models spend far more tokens per request on longer-horizon, agentic tasks, so a flat request count no longer maps to a flat cost. A single request can now burn through tokens that an older one never would.
Two wording choices turned a reasonable change into a backlash. First, Cursor described the included pool in terms of “rate limits,” which read like throttling rather than a spend cap. Second, the word “unlimited” was attached to usage without making clear it applied only to Auto, the setting that routes a task to whichever frontier model has capacity, and not to a model the user pins by hand. Cursor later acknowledged both directly: in its words, “we were not clear that unlimited usage was only for Auto and not all other models,” and it offered refunds for surprise charges run up between June 16 and July 4, 2025.
The tiers, in dollars
Every paid Cursor plan is a monthly credit pool denominated in dollars, plus the editor itself. The pools, as listed on Cursor’s pricing page and compiled across current pricing guides:
| Plan | Price / month | Included usage pool | Notes |
|---|---|---|---|
| Hobby | $0 | Limited Agent + Tab | Free, no card required |
| Pro | $20 | ~$20 of API-rate usage | The default paid tier |
| Pro+ | $60 | ~$60 (about 3x Pro) | For heavier solo use |
| Ultra | $200 | ~$400 (about 20x Pro) | Only tier where usage value exceeds the price |
| Teams | $40 / user | Pooled across the team | Adds admin, SSO, privacy mode |
Two details are worth pulling out. Annual billing takes 20% off every paid tier, so Pro effectively runs about $16/mo on a yearly commitment. And Ultra is the only plan where the included usage is worth more than the sticker: roughly $400 of API-rate usage for $200, against Pro and Pro+ where the pool roughly equals the price. For a genuinely heavy user, Ultra is the one tier that is cheaper than paying for the same usage à la carte. The Pro+ and Ultra multipliers come from pricing guides rather than the headline page, so treat them as approximate.
What $20 actually buys
Here is the number people search for, and the reason it cannot be one number. Cursor’s own clarification gives worked examples for the $20 Pro pool: roughly 225 requests on Claude Sonnet, about 550 on Gemini, and about 650 on a GPT 4.1-class model.
| Item | Value |
|---|---|
| GPT 4.1-class | ~650 |
| Gemini | ~550 |
| Claude Sonnet | ~225 |
Run the division and the mechanism is obvious. At about 225 requests, a Sonnet request costs the pool roughly $0.089. At about 650, a GPT 4.1-class request costs roughly $0.031. That is nearly a 3x spread on the price of “one request,” which is exactly why a fixed request count had to go: a request stopped being a fixed unit of cost. The honest unit for any agentic tool is cost to finish the work, not cost per request or per token, and Cursor’s pricing is now built around that reality whether or not the framing landed well.
The practical read: if the work leans on Claude Sonnet or another premium model, the $20 pool drains roughly three times faster than the cheaper-model headline suggests. Plan the tier around the model, not the other way around.
When the bill goes past $20
The pool is not a hard wall by default. When a Pro user spends the $20, Cursor applies a short grace period, then switches to on-demand usage billed in arrears at API rates. That is the line item that surprises people: the editor keeps working past the included pool, and the overage shows up later.
There are two levers. Set a spend limit, and Cursor stops at the cap instead of billing on-demand. Or lean on Auto, which routes to available frontier models and does not draw down the pinned-model pool the same way, so it is the closest thing to the old “unlimited” feel. The mistake to avoid is pinning a premium model by name, ignoring the pool meter, and discovering the overage on the next invoice.
Is Cursor Pro worth $20?
For most individual developers, yes, on one condition: the work fits inside the $20 pool, or the spend cap is set where the budget can absorb the overage. Pro buys the full editor, frontier-model access, and a predictable floor. The risk is not the $20. It is the on-demand spend above it when a premium model is pinned and left unwatched.
The decision is really a usage-versus-flat-fee question, and that is worth modeling rather than guessing. Compare the subscription pool against paying the same model usage straight through an API in the subscription vs API calculator, and size the per-task cost in the coding-agent cost calculator. For where Cursor sits against the rest of the field on raw price, the AI coding agent pricing pillar and the coding-agents hub lay out the full ladder.
Two adjacent pieces round out the picture. For the cost of Cursor’s in-house model specifically, see the breakdown of what Composer 2.5 actually costs per task. And for what new ownership could mean for future pricing, SpaceX’s acquisition of Cursor and the cost of leaving covers the lock-in math.
The bottom line
Cursor pricing is not as confusing as the backlash made it sound, once the unit is right. The plans sell dollars of model usage, not requests. The $20 Pro pool is real spending power at API rates, worth about 225 premium-model requests or roughly 650 on a cheaper model. The overage is opt-out, not opt-in, so the spend limit is the setting that matters. Pick the tier around the model the work actually uses, cap the spend, and the pricing stops being a mystery.
Frequently asked questions
- How many requests does Cursor Pro give you?
- There is no fixed request count anymore. Pro is a $20 monthly pool of model usage charged at API rates. Cursor's own examples put that at about 225 Claude Sonnet requests, about 550 Gemini requests, or about 650 on a GPT 4.1-class model. The count depends on the model and how token-heavy each task is.
- What happened to Cursor pricing?
- In 2025 Cursor moved from a fixed request count (widely understood as 500 requests on Pro) to a usage-credit model: a $20 monthly pool spent at API rates. Newer agentic models burn far more tokens per request, so a flat request count no longer mapped to a flat cost. Confusing "rate limits" language and an "unlimited" label that applied only to Auto routing drove most of the backlash.
- How much does Cursor cost in 2026?
- Hobby is free. Pro is $20/mo, Pro+ is $60/mo, and Ultra is $200/mo, each a usage pool of that dollar value (Ultra includes roughly $400 of API-rate usage). Teams is $40 per user per month. Annual billing takes 20% off every paid tier.
- What happens when you hit the Cursor usage limit?
- When the included pool is spent, Cursor applies a short grace period, then bills on-demand usage in arrears at API rates unless a spend limit is set to stop it. Auto routing keeps working without drawing down a pinned model the same way, so it is the closest thing to the old unlimited behavior.
- Is Cursor Pro worth $20 a month?
- For most individual developers, yes, provided the work fits the $20 pool or a spend cap is set. The full editor and frontier-model access are worth the floor; the risk is on-demand overage above $20 when a premium model is pinned and left unwatched. Model it against API pricing before committing to a tier.
Sources
- Cursor (Anysphere). (2025). Clarifying our pricing [vendor blog; the $20 Pro usage pool, the ~225 Sonnet / ~550 Gemini / ~650 GPT 4.1 request examples, the Auto-only “unlimited” acknowledgement, and the June 16 to July 4 2025 refund window]. cursor.com/blog/june-2025-pricing
- Cursor (Anysphere). (2026). Pricing [vendor pricing page; Hobby free, Pro $20/mo, Pro+ and Ultra tiers, Teams $40/user, annual discount]. cursor.com/pricing
- No Code MBA. (2026). Cursor Pricing 2026: All 6 Plans ($0 to $200/mo) & Hidden Costs [secondary; Pro+ $60 and Ultra $200 tier prices and the approximate 3x / 20x usage multipliers]. nocode.mba/articles/cursor-pricing