Cursor’s pricing history has three main eras. Early paid plans counted premium requests, which made capacity easy to state but treated cheap and expensive Agent work alike. On 16 June 2025 Cursor moved individual Pro toward $20 of model inference at API prices, made Auto unlimited at the time, and launched Ultra at a historical $200 monthly price. After unclear communication and unexpected charges, Cursor apologized on 4 July and offered refunds for affected overage. Teams moved from fixed Sonnet requests to API-valued usage later in 2025. In June 2026 Cursor introduced separate first-party and third-party pools for Teams and a Premium seat; July added Router optimization modes and the India-only Start plan. Current plan mechanics should always come from live pricing and account pages.
- Cursor pricing changed units, not only dollar amounts: requests, compute value, and then separate model pools.
- The 16 June 2025 individual change replaced the default 500-request method with at least $20 of API-priced inference on Pro.
- Cursor’s 4 July 2025 apology acknowledged unclear communication and offered refunds for unexpected usage charges.
- Teams moved from fixed Sonnet requests to API-based usage plus a token surcharge later in 2025.
- The June 2026 Teams update created first-party and third-party pools and a 5x Premium seat.
- For every future change, compare unit, allowance, reset, overage, routing, and migration date, not the headline price alone.
The timeline at a glance
Historical terms are preserved as historical terms. Current buying decisions belong to the live plan picker.
| Date or era | What Cursor sold | What it meant for users |
|---|---|---|
| Before 16 June 2025 | Pro counted premium requests; the later clarification identifies a 500-request monthly limit, with some Sonnet requests costing two | Capacity looked simple, but a tiny question and a long Agent task consumed coarse units |
| 16 June 2025 | Pro moved by default to compute limits with at least $20 of inference at API prices; Auto became unlimited at the time; Ultra launched at a historical $200/mo with 20x Pro usage | Model and token choice began deciding how long the same $20 seat lasted |
| 30 June 2025 | Cursor revised the announcement and pricing-page wording | The company was still clarifying what “unlimited” applied to |
| 4 July 2025 | Public apology, detailed explanation, and refund offer for unexpected charges from 16 June through 4 July | The billing mechanics became clearer, but trust became part of the price |
| 12 August 2025 announcement | Teams scheduled a move from 250 Sonnet requests plus $0.08 overage to $20 of API-priced Agent usage plus $0.25 per million total tokens | Teams inherited model-sensitive cost and a new platform surcharge |
| 1 June 2026 announcement | Teams got separate first-party and third-party pools; Standard stayed $40 monthly; Premium launched with 5x usage | Cheap first-party capacity was separated from expensive third-party inference |
| 22 July 2026 | Cursor Router split Auto into Cost, Balance, and Intelligence optimization choices | “Use Auto” stopped being complete pricing advice; the optimization mode matters |
| 28 July 2026 | Start launched in India at ₹649 monthly, focused on Cursor models without a third-party pool | Cursor added regional pricing and a first-party-only economic shape |
| August 2026 snapshot | Hobby, Pro, Pro Plus, Ultra, Start, Teams Standard and Premium, and Enterprise, with model pools and optional on demand | The account dashboard and current plan picker are now more authoritative than request counts in articles |
The dates explain why two apparently reputable pages can disagree. A 2024 guide may accurately describe free completions and slow premium requests. A July 2025 article may accurately say Auto is unlimited and Ultra costs $200. An August 2026 account can present different routes, inclusions, and signed-in prices. Accuracy decays when a page omits the date and keeps the confident number.
Era one: fast and slow requests
The early pricing model sold an intuitive monthly count. Users discussed fast premium requests, slow requests after the fast allocation, and fixed free-tier completion or small-model counts. The July 2025 clarification later summarized the individual Pro baseline as 500 external-model requests a month, with Sonnet models costing two requests. That is the origin of the “how many Cursor requests do I get?” search language that survives today.
| What request pricing did well | What it did badly |
|---|---|
| Made the allowance easy to state | Pretended requests had similar underlying cost |
| Let users count down without token arithmetic | Penalized a short prompt and subsidized a huge Agent run within the same coarse unit |
| Created a visible slow fallback in some plan generations | Made model multipliers and fast-versus-slow rules another layer to memorize |
| Kept the seat bill predictable | Became harder for Cursor to sustain as agents used longer context and more tools |
The unit broke as the product became more agentic. One completion might ask for a handful of tokens. One Agent turn could read files, run a test, inspect output, and continue through several tool calls. Cursor later said difficult requests could cost an order of magnitude more than simple ones even on the same model. A fixed request price therefore overcharged small interactions or undercharged large ones.
16 June 2025: requests became compute
Cursor announced two linked individual-plan changes on 16 June 2025. Pro would move by default from request limits to compute limits, with at least $20 of model inference at API prices each month. Auto would be unlimited at that time and tool-call limits would be lifted. Existing users could initially choose to stay on the 500-request method. Cursor also introduced Ultra at a historical launch price of $200 a month with twenty times more usage than Pro.
The sticker price of Pro stayed at $20. The object bought for $20 changed. Instead of a fixed number of coarse requests, users received a dollar-valued pool whose practical request count depended on the chosen model and the work inside each turn. That could be more generous for cheap, short requests and less generous for expensive, long-context models.
| Before | After the June announcement | User consequence |
|---|---|---|
| 500-request method | $20 or more of API-priced model inference | No single reliable request count |
| Model multiplier against requests | Model rate and tokens against dollars | Prompt shape and context directly affected capacity |
| Tool calls constrained as part of request behavior | Tool-call limits lifted | Longer agents could act more freely but still consume model inference |
| No Ultra tier in the announcement history | Ultra at $200/mo, 20x Pro usage at launch | Power users could buy predictability instead of open-ended overage |
| Auto priced like premium models before June 2025 | Auto announced as unlimited for individuals | Routing became the escape valve after the frontier pool |
Cursor's rationale was economically coherent: token-valued inference tracks provider cost better than request counting. The failure was communication. “Unlimited” appeared beside a plan that still had a limited frontier-model pool, and users did not consistently understand which requests would continue free, which would consume included usage, and which settings could generate additional charges.
4 July 2025: the apology and refunds
On 4 July Cursor published “Clarifying our pricing.” It said the changes had not been communicated clearly, accepted responsibility, and offered refunds for unexpected usage charges incurred between 16 June and 4 July. The post laid out its own mini-timeline: original announcement on 16 June, revised wording on 30 June, and the apology on 4 July.
The clarification said new Pro provided unlimited Tab and Auto at that time, $20 of frontier-model usage each month at API pricing, a short grace period after the pool, and the option to enable a spend limit for more usage at cost. It also translated the pool into model-specific median request estimates. Those estimates were a snapshot, not a new contractual request allowance.
| Complaint underneath the outrage | Why it mattered |
|---|---|
| “Unlimited” did not mean every model | The headline described Auto while users inferred the whole product |
| The unit changed under a familiar $20 price | A user’s historic request volume no longer predicted capacity |
| Unexpected overage appeared | A productivity tool crossed from fixed seat to variable bill |
| Documentation changed during rollout | Users comparing screenshots were often reading different wording |
| Request estimates looked like guarantees | Median examples vary by model, context, and future rates |
Cursor also promised better advance notice, clearer documentation, support readiness, and more dashboard visibility. The lasting product response was not only copy. Usage visibility became part of later pricing announcements, including editor notifications, dashboard breakdowns, and team alerts. When a pricing system depends on invisible routing and variable token cost, observability is part of the product rather than an optional analytics extra.
Teams moved to API-priced usage
Individual plans changed first. On 12 August 2025 Cursor announced that Teams would move at each renewal after 15 September from fixed-cost request pricing to API-based usage. The old Team structure in that announcement included unlimited Tab, 250 Sonnet requests each month, and additional Sonnet usage at $0.08 per request.
The replacement included unlimited Tab, $20 of Agent usage per user, and the ability to buy more. Third-party usage would be valued at the public model API price plus a Cursor Token Rate of $0.25 per million total tokens. Cursor said the extra charge covered code indexing, indexing-model cost, and tool execution with custom models.
For administrators, the change traded an easy per-request forecast for a closer match to provider economics. The same team activity could now cost differently depending on model selection and task shape. That made model policy, user-level limits, and a reliable admin dashboard materially more important. It also exposed a conceptual difference between paying for model inference and paying Cursor's platform layer.
The migration date mattered. A team before renewal and a team after renewal could truthfully report different meters at the same time. Any historical investigation should record the account renewal date, not only the announcement date. Pricing changes often roll out by billing cycle rather than globally at publication time.
July 2026: Auto became three economic choices
Cursor Router arrived on 22 July 2026 and made old “just use Auto” advice incomplete. Auto gained Cost, Balance, and Intelligence optimization modes. Cost seeks price-efficient intelligence. Balance routes toward strong daily-driver models. Intelligence favors frontier quality. Cursor's announcement says Balance and Intelligence bill at the routed model's rate; the current site pricing guide describes Auto Cost with a flatter per-token treatment.
| Router mode | Optimization goal | Budget implication |
|---|---|---|
| Cost | Good quality at efficient price | The predictable default when preserving allowance matters |
| Balance | Strong daily quality | Model-dependent cost based on the route selected |
| Intelligence | Highest available capability | Can select expensive frontier routes and drain allowance faster |
The name Auto survived while its pricing behavior changed. That is exactly the kind of label reuse that makes historical advice dangerous. In June 2025 “Auto is unlimited” was an important escape hatch for individuals. In the August 2026 site snapshot, the Router mode underneath Auto determines how requests are priced and which pool moves. Always state the date and optimization mode.
On 28 July Cursor added Start in India at ₹649 per month, tax inclusive, with UPI or card payment. The plan centers on generous access to Cursor models, cloud agents, mobile access, and extensibility, without a third-party model pool or on-demand route in the site snapshot. This was both regional pricing and product segmentation by model economics: a lower local price made possible by keeping usage inside first-party routes.
What the current structure means
As of the August 2026 site snapshot, Hobby is free, Pro is $20 monthly, Teams Standard is $40 per user monthly, and Enterprise is custom. The site pricing guides describe Pro Plus and Ultra by capacity and advise checking the signed-in plan picker for current prices. Paid individuals receive unlimited Tab, generous first-party model usage, and a third-party allowance: $20 on Pro, $70 on Pro Plus, and $400 on Ultra.
| Current component | Stable interpretation |
|---|---|
| Seat price | Fixed subscription charge for the plan period |
| Cursor Models pool | Included first-party capacity, described as generous rather than one universal request count |
| Other Models pool | Dollar-valued third-party inference by model rate |
| Tab | Unlimited on paid individual plans |
| On demand | Optional continuation after included usage, billed in arrears at model rates |
| Hard spend limit | The boundary that keeps optional continuation from becoming an unlimited invoice |
| Monthly reset | Included usage refreshes on the account billing cycle, including annual subscriptions |
Current does not mean permanent. Cursor changes models, routes, plan copy, and dashboard presentation frequently. The durable mental model is to identify which party supplies the model, which pool pays for it, what happens when that pool empties, and whether the next request stops or becomes billable.
How to judge the next Cursor pricing change
The next pricing cycle will probably arrive with a new model, plan name, routing promise, or usage multiplier. Ignore the announcement adjectives and fill in this table. If any row is missing, the economics are not yet clear enough to compare.
| Question | Why it matters |
|---|---|
| What is the unit now? | Requests, tokens, dollars, compute, and model pools cannot be compared directly |
| What exact allowance is guaranteed? | Bonus or “generous” capacity may not be contractual or uniform |
| Which models draw which pool? | The picker can change the bill without changing visible workload |
| What happens at zero? | Stop, route, degrade, queue, or bill are radically different failure modes |
| Is overage opt-in and capped? | A fixed seat can quietly become variable spend |
| When does it reset? | Monthly billing date, rolling window, and calendar month produce different planning |
| Who migrates, and when? | New users and renewing customers may coexist on different rules |
| What historical users can keep? | Grandfathering can make another user’s screenshot irrelevant |
| What visibility exists before the limit? | A variable system without alerts transfers surprise to the user |
| What is the live price in my account? | Regional, annual, seat, and signed-in prices can differ from public summaries |
Archive the before state
Save the current plan name, renewal date, included pools, model default, on-demand cap, and last full-cycle usage. Without a before state, nobody can prove whether value improved.
Normalize one real month
Reprice the last completed cycle under the new rules using actual model and pool usage. Do not compare marketing request equivalents.
Check the failure mode
A lower theoretical cost can still be worse if it turns a hard stop into silent overage, or vice versa. Choose based on the risk your workflow can tolerate.
Wait for account-specific migration details
The blog gives the product direction. The billing portal and renewal notice determine what happens to this account and when.
Measure the first full cycle
Track by pool, model, and task shape. One launch-week spike is not a monthly baseline, and a partial cycle cannot be compared honestly with a full one.
Questions people ask
The major individual change was announced on 16 June 2025, revised on 30 June, and clarified with an apology on 4 July. Teams announced an API-priced transition on 12 August 2025 and a two-pool structure on 1 June 2026.
Pro moved from request counting to $20 of API-priced model inference while Auto was described as unlimited. Cursor later said communication was unclear, apologized, and offered refunds for unexpected usage charges incurred from 16 June through 4 July 2025.
The historical Pro method counted 500 external-model requests per month, with some Sonnet usage costing two requests. Other older plan generations also discussed fast and slow premium requests.
Pro moved by default from request limits to compute limits with at least $20 of model inference at API rates, Auto became unlimited at that time, tool-call limits were lifted, and Ultra launched historically at $200 monthly with 20x Pro usage.
Cursor’s 4 July 2025 post offered refunds for unexpected usage charges incurred between 16 June and 4 July. That was a historical remediation window, not a standing refund policy.
Teams first moved from fixed Sonnet requests to API-valued Agent usage after September 2025 renewals. In June 2026 Cursor announced separate first-party and third-party pools plus a Premium seat with 5x Standard usage.
Do not carry the June 2025 rule forward. By August 2026 Auto uses Cursor Router with Cost, Balance, and Intelligence modes, and the pricing behavior depends on the mode and routed model.
They often describe different eras, mix historical launch prices with current allowances, or turn model-specific median examples into fixed request limits. Check the article date and the live signed-in plan picker.
Compare the unit, guaranteed allowance, model-to-pool routing, reset timing, overage behavior, hard caps, migration date, and your own last full month under both systems.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.