Since 12 October 2026 there is no $20 flat Cursor plan to buy. New buyers choose Plus at $8 — Grok and Composer only, no third-party models, no agent surfaces — or Premium at $30, where Claude and GPT are still in the editor but billed at API list price from the first request. At a moderate agentic workload that is about $76 a month instead of $20. The picks: Grok-first, take Plus at $8, the cheapest credible setup in the category; Claude-first, take Claude Pro at $20; GPT-first, take Codex at $20; mixed models in an IDE, take Copilot Pro+ at $39; and if you hold legacy Cursor Pro, stay — switching is permanent and the docs say your plan continues indefinitely.
What actually changed, in one paragraph
SpaceXAI merged Cursor, Grok and X Premium into one ladder — Free, Plus $8, Premium $30, Super $100, Ultra $200, Teams from $30 per seat, Enterprise custom — sold through X Subscription Corp. Read as a price list it is a cut. Read as a billing model it is a reclassification: the included weekly pool covers Cursor’s own models (Grok 4.5, 4.6, 4.7 and Composer 2.5), and Claude, GPT and Gemini moved outside it onto pass-through API billing from prepaid credit. The subscription fee no longer predicts the bill for the majority of Cursor users, because Auto mode and the third-party models were what most people ran.
One correction worth carrying, since most coverage including our own first pass got it wrong: the pool is not Grok-only. Cursor’s docs say “the pool covers Grok 4.7, Grok 4.6, Grok 4.5, and Composer 2.5,” and Composer is Cursor’s own model, priced on the same page at $0.50 input / $0.20 cached / $2.50 output per million. Plus is the exception — its plan card says “only Grok models are available on this plan,” which does not obviously reconcile with the pool description, and Cursor has not explained the gap.
The arithmetic nobody publishes
Everything below uses one workload so the numbers are comparable: a developer running Claude Sonnet 5.5 at $2.00 input / $0.10 cached input / $10.00 output per million tokens, over 21 working days. Rates verified 21 October 2026 against Cursor’s models and pricing page; plan prices against cursor.com/pricing, claude.com/pricing, learn.chatgpt.com and github.com/features/copilot/plans. The pricing tracker carries the dated before-and-after and the API cost calculator will re-run this for your own token mix.
| Daily agentic load | Cached prefix | Fresh input | Output | Tokens/day | API cost/day | + Premium $30 |
|---|---|---|---|---|---|---|
| Light (~1h of agent work) | 1M | 70K | 50K | 1.12M | $0.74 | $45.54/mo |
| Moderate (~3h) | 3M | 200K | 150K | 3.35M | $2.20 | $76.20/mo |
| Heavy (all day) | 9M | 600K | 450K | 10.05M | $6.60 | $168.60/mo |
Two structural facts the table makes visible. First, the light user is the one who loses most in proportion: $45.54 against $20 is a 128% rise for someone whose usage would have been comfortably inside the old Pro plan. The flat fee was a subsidy flowing from heavy users to light ones, and removing it bills the light user for a pool they never drained. Second, cache reads dominate the token count and not the cost — 90% of tokens, 14% of spend at moderate load — which is fine under Anthropic’s rate card and becomes the whole problem on Teams.
Teams: where the surcharge inverts
Cursor bills Teams and Enterprise third-party requests at API list plus a Cursor Token Rate of $0.25 per million tokens, applied to input, output and cached tokens. Twenty developers at moderate load:
| Line | Legacy Teams Standard | New Teams Premium seat |
|---|---|---|
| Seats | 20 × $40 = $800 | 20 × $30 = $600 |
| Third-party model tokens | Included via the “Other Models” pool | 20 × $46.20 = $924 |
| Cursor Token Rate | — | 1,407M × $0.25 = $352 |
| Monthly total | $800 | $1,876 |
| Per developer | $40 | $93.80 |
The seat price fell 25% and the bill rose 135%. And the surcharge line is where the cache arithmetic turns perverse: of those 1,407 million tokens, 1,260 million are cache reads that Anthropic prices at $0.10 per million and Cursor surcharges at $0.25 per million. On the cheapest and most numerous token in agentic work, the platform fee is 2.5x the model fee. Cursor’s first-party models are exempt from the rate, which tells you where the pricing is steering a team — and a team that follows the steering has bought a single-model editor.
Two honest caveats. The $0.25 figure and its application to cached tokens come from the tail of Cursor’s models-and-pricing page, which resisted full retrieval on 21 October; the plan table’s “plus the Cursor Token Rate on eligible models” wording is verbatim, the cached-token detail is not. And an exclusion list reportedly covers Claude Haiku 5.5, several Gemini versions and GPT-4o Mini, so a team that deliberately runs a small model escapes the surcharge entirely. Also note Cursor’s documented 10% uplift for regional or US-only data residency on eligible models, which stacks on top of everything above.
The shortlist, metered against flat
| Option | Cost / dev / month | Model billing | Best when |
|---|---|---|---|
| Cursor Plus | $8 ($80/yr) | Flat — pool covers Grok 4.x + Composer | Grok suits your code and you never mix models |
| Cursor Premium | $30 + metered ($300/yr) | Hybrid — third-party at API list | You need Cursor’s agent surfaces specifically |
| Legacy Cursor Pro | $20 | Flat — included “Other Models” pool | You already hold it. Do not switch |
| Claude Code | $20 (Pro) → $100/$200 (Max) | Flat, with rolling usage windows | Claude-first, terminal-comfortable |
| Codex | $20 (Plus) → $100/$200/$500 (Pro) | Flat, with purchasable credits | GPT-first; also the only first-party GPT path |
| Copilot Pro | $10 + overage at par | $15 of credits included | Light load, mixed models |
| Copilot Pro+ | $39 | $70 of credits included | Moderate load, mixed models, IDE-shaped |
| Copilot Max | $100 | $200 of credits included | Heavy agentic load, predictable ceiling |
| Devin Desktop (ex-Windsurf) | $20 (Pro), $200 (Max) | Flat | You want a Cursor-shaped editor elsewhere |
| OpenCode / Aider | Free + your API spend | Fully metered, BYOK | You are in a terminal and want no platform fee |
Read the Copilot rows against the Cursor rows and the real difference is legibility rather than price. Copilot publishes included spend in dollars, so $39 buying $70 of tokens is a number you can check against your own usage before you pay. Cursor publishes 4x, 15x and 50x of a Plus allowance whose absolute size appears nowhere, which means no Cursor buyer can compute their effective rate in advance — the only verifiable figures in a Cursor bill are the subscription fee and the list rates for models outside the pool, and the second is not a cap. For a side-by-side of every published allowance in the category, see best AI coding tools.
Where the flat-fee options bite back is capacity, and it is worth saying plainly rather than burying: Claude Pro and ChatGPT Plus cap you with rolling usage windows, so “flat” means your bill is predictable and your throughput is not. The heavy-load developer who genuinely runs ten million tokens a day will hit those windows, and for that profile Copilot Max at $100 for $200 of credits or Claude Max at $100 is the honest comparison — not Premium plus $169 of metered tokens, but not $20 either.
Decision table
| If you… | Buy | Why |
|---|---|---|
| Hold legacy Pro or legacy Teams | Nothing. Stay | Switching is permanent and drops the included third-party allowance |
| Run Grok or Composer by choice | Cursor Plus, $8 | Cheapest credible setup in the category, bundled apps included |
| Run Claude all day, terminal is fine | Claude Pro, $20 | Model included; Cursor vs Claude Code covers the editor trade |
| Run GPT and need it to keep working | ChatGPT Plus, $20 | First-party, and not downstream of the 12 November cutoff |
| Mix models and want an IDE | Copilot Pro+, $39 | Only tool here publishing included spend in dollars; see Copilot vs Cursor |
| Run heavy agentic loads | Copilot Max $100 or Claude Max $100 | Metered Premium costs more and caps nothing |
| Want zero platform fee | OpenCode or Aider | Pure API list, any provider, no surcharge |
| Are a team on legacy seats | Stay, and model the surcharge first | $40/seat including a pool against $93.80/seat metered |
What this is really a lesson in
The thing to take from the repricing is not that Cursor got dearer. It is that a flat fee on top of metered inputs is a promise, not a cost structure, and promises get revisited when the owner changes. Cursor spent two years selling one number that covered every model, which only worked while the company was willing to absorb the spread between what users burned and what they paid. SpaceX’s acquisition gave it both a reason and a first-party model to steer toward, and the reclassification followed within two months — alongside OpenAI terminating supply and annual billing appearing in the docs only.
The practical test to apply to any tool in this category, before you buy: can you divide the subscription fee by something to get a rate? Copilot passes — credits are denominated in cents. Claude and ChatGPT pass differently, by capping throughput instead of billing it. Cursor’s new ladder does not pass at any tier, and the reason it does not is that the absolute size of a Plus allowance is unpublished. A plan you cannot price is a plan whose price can move without an announcement, which is the position every Cursor Premium buyer is now in.
What changed
- 21 Oct 2026: Page created. Corrected the widely-repeated claim that Cursor’s new pool is Grok-only — the docs say it covers Grok 4.5/4.6/4.7 and Composer 2.5. Confirmed from Cursor’s legacy-plans page that switching is “optional and permanent” and legacy plans “continue unless you change or cancel”, so no migration date exists rather than merely being unannounced.
- 15 Oct 2026: Cursor annual billing appeared in the billing docs only — $80/$300/$1,000/$2,000 a year, two months free, individual tiers only.
- 12 Oct 2026: The X Subscription replaced the Cursor ladder. Pro at $20 closed to new signups; third-party models left the included pool.
- 12 Nov 2026 (scheduled): proposed end of OpenAI model supply to Cursor — a ceiling rather than a deadline. See the migration guide.
- 14 Aug 2026: SpaceX’s $60B acquisition of Anysphere closed, opening the change-of-control window.
Frequently asked questions
Is Cursor actually more expensive now?
It depends entirely on which model you run, and the two answers point in opposite directions. If your daily driver is Grok or Cursor's own Composer, Cursor got cheaper: Plus at $8 a month includes those models in the usage pool, against $20 for the old Pro plan, and it bundles Grok Chat, Voice, Imagine, the Grok API and X Premium features on top. If your daily driver is Claude or GPT, Cursor got substantially dearer, because those models left the included pool entirely. Cursor's docs state that third-party models "aren't part of the included pool" and that on Premium and above "they're available in Cursor at the model's API list price from the first request, paid from your prepaid credit." There is no free allowance to burn through first. At a moderate agentic workload — roughly 3M cached-prefix tokens, 200K fresh input and 150K output a day on Claude Sonnet 5.5 — that is about $46 a month of tokens on top of Premium's $30, so about $76 against the $20 a legacy Pro subscriber still pays.
I hold legacy Cursor Pro at $20. Should I switch or stay?
Stay, and the documentation is unusually clear about why. Cursor's legacy-plans page states that "switching to The X Subscription is optional and permanent" and that "your legacy plan, price, and features continue unless you change or cancel." It also warns in its own words that each new tier "drops the included third-party model allowance" and that on switching "third-party models stop being included." That is a one-way door out of a grandfathered rate, with no published date forcing you through it. The suggested mapping is Pro to Premium, Pro+ to Super, Ultra to Ultra, Teams Standard to a Teams Premium seat and Teams Premium to a Teams Super seat — note that Pro to Premium is $20 to $30 before a single metered token. The one reason to switch voluntarily is if you want the bundled Grok apps and the X subscription, and your editor work is already Grok-first. Switching also cancels eligible subscriptions with a prorated refund and issues one final invoice for uninvoiced on-demand usage, so time it at a billing boundary.
What is the cheapest credible setup in the category right now?
Cursor Plus at $8 a month, provided you are willing to run Grok 4.7 and Composer 2.5 rather than Claude or GPT. Nothing else on the shortlist is close on price, and it is not a stripped tier in the usual sense — the pool covers Agent in the editor, the CLI, cloud agents and the Grok apps. The catch is in the plan card's own small print: Plus has "no Auto, Cloud Agents, Automations, Bugbot, Origin, third-party models, or your own API keys." So $8 buys you the editor and one model family with no escape hatch, which is a reasonable trade if Grok suits your code and a bad one if you expect to mix models. If you want the $8 price and the bundled apps but also need the agent surfaces, the next step is Premium at $30, and at that point you are comparing against flat-fee tools that include their model.
Which alternative includes the model in the subscription?
Claude Pro at $20, ChatGPT Plus at $20 and GitHub Copilot's paid plans all bill the model inside the fee rather than beside it, though they do it in two different ways. Claude Pro and ChatGPT Plus are genuinely flat: you pay $20, you run the model, and when you hit the rolling usage window you wait or buy credits — the bill cannot surprise you. Copilot is a hybrid that is more honest than it looks: every plan includes a dollar-denominated allowance of AI credits at one cent each, so Pro at $10 includes $15 of token spend, Pro+ at $39 includes $70 and Max at $100 includes $200, with overage at par. That makes Copilot the only tool here where you can divide included spend by seat price and get a real number — $1.50, $1.79 and $2.00 of tokens per dollar respectively — which is exactly the comparison Cursor's 4x/15x/50x multiples make impossible.
Does the Cursor Token Rate apply to me?
Only on Teams and Enterprise, and there it deserves arithmetic rather than a shrug. Cursor's plan table bills Teams and Enterprise third-party requests at "API list price plus the Cursor Token Rate on eligible models," documented at $0.25 per million tokens and applied to input, output and cached tokens alike, with an exclusion list that reportedly covers Claude Haiku 5.5, several Gemini versions and GPT-4o Mini. Individual Premium, Super and Ultra plans are shown as plain API list price from prepaid credit, with no surcharge. The reason it matters more than the figure suggests is cache reads. On a typical agentic day, the large majority of tokens are cached prefix — and Claude Sonnet 5.5 charges $0.10 per million for a cache read while Cursor's surcharge on the same token is $0.25. On the cheapest and most numerous token in agentic work, the platform fee is two and a half times the model fee.
Does the 12 November OpenAI cutoff change this?
It narrows the choice for one group specifically: anyone whose reason for paying Cursor was GPT. OpenAI moved to end the contract that supplies its models to Cursor with a proposed shutoff of 12 November 2026, which is the latest date its change-of-control window allows rather than a target, and a personal API key restores local Chat and Agent only — not Tab, Auto mode, cloud agents, Automations or the CLI. So a new buyer choosing Premium at $30 in order to run metered GPT is buying a metered path to a model that may stop being available within weeks, on the surfaces that matter least. If GPT is the requirement, Codex inside ChatGPT Plus at $20 is both cheaper and first-party. Our migration guide covers the feature-by-feature survival table.