Cursor Models: Which Model Should You Use?

Cursor sorts its models by how you pay for them. Its own models (Grok 4.7Cursor and SpaceXAI's current lead Cursor model, in the weekly pool at $2 per million input tokens and $6 per million output, with a 256k context window and a 500k variant; above 256k input tokens it bills at 2x the standard rate. Press Enter for the full definition., Grok 4.6A Cursor and SpaceXAI model for complex coding and knowledge work, still in the Cursor model pool and succeeded by Grok 4.7; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition., Grok 4.5 and Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition.) draw from the weekly pool included with every paid plan on The X SubscriptionThe subscription Cursor is sold through: one plan for Grok, Cursor and X, with Free, Plus, Premium, Super and Ultra tiers, per-seat Teams, a weekly included pool for Cursor models shared with Grok, and third-party models billed at API list price from prepaid credit. Press Enter for the full definition.. Third-party models from Anthropic, Google, OpenAI and Meta sit outside that pool and bill at the model's API list price from prepaid credit, on Premium and above. Free and Plus don't include them. Legacy Cursor plansThe plans Cursor sold before The X Subscription (Hobby, Start, Pro, Pro+, Ultra, and Teams Standard and Premium seats); they keep working for existing subscribers with two monthly usage pools and on-demand usage. Press Enter for the full definition. make the same split with two monthly pools, Cursor ModelsOn legacy Cursor plans, the monthly usage pool for Cursor's own models (Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5), kept separate from the third-party Other Models pool. The X Subscription replaces it with one weekly pool. Press Enter for the full definition. and Other ModelsOn legacy Cursor plans, the monthly pool for supported third-party models from providers such as Anthropic, Google and OpenAI, separate from Cursor Models. On The X Subscription third-party models are billed from prepaid credit instead. Press Enter for the full definition..

Grok 4.7Cursor and SpaceXAI's current lead Cursor model, in the weekly pool at $2 per million input tokens and $6 per million output, with a 256k context window and a 500k variant; above 256k input tokens it bills at 2x the standard rate. Press Enter for the full definition. is now Cursor's lead model. Alongside it, Cursor lists a newer generation of third-party models: Claude Opus 5.5Anthropic's top Opus-tier model in Cursor at $4 per million input tokens and $20 per million output, 20% below Claude Opus 5, which it replaces. Press Enter for the full definition., Sonnet 5.5, Haiku 5.5 and Fable 5.1, Gemini 3.8 FlashGoogle's current speed-tier model in Cursor at $0.75 per million input tokens and $3.75 per million output; it succeeds Gemini 3.7 Flash. Press Enter for the full definition., Gemini 4 ArgonGoogle's newest Gemini model in Cursor at $2 per million input tokens and $10 per million output. Press Enter for the full definition. and Meta's Muse Wonder. Cursor's pricing table marks Claude Opus 5, Fable 5 and Sonnet 5, Gemini 3.7 FlashGoogle's previous speed-tier model in Cursor, hidden by default and succeeded by Gemini 3.8 Flash at the same published rates. Press Enter for the full definition. and Gemini 3.6 Flash as hidden by default, so this guide treats them as the previous generation.

Grok 4.7Cursor and SpaceXAI's current lead Cursor model, in the weekly pool at $2 per million input tokens and $6 per million output, with a 256k context window and a 500k variant; above 256k input tokens it bills at 2x the standard rate. Press Enter for the full definition. is a model you select for a request. Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition. is a separate product: a persistent teammate with apps, routines, a shared cloud computer and its own eligibility rules. The shared name doesn't make their setup, data boundaries or billing interchangeable.

Previous generation, hidden by default

Cursor still documents these models and names a successor for each, but it hides them by default. Pin one only when you need a controlled comparison with its successor, or when an older approval names it.

Pick the least expensive model that can finish the task cleanly, which is not always the model with the lowest rate. A weak first attempt adds review work, fills the conversation with a wrong path and often forces a second run. The table below is our starting policy, not a benchmark result.

A clear task boundary does more for the result than the model name. Give the model the relevant files, the expected behavior and one check it can run. Then judge the diff and the check results instead of the answer's confidence. A model that finishes in one pass on your repository often costs less than a lower-rate model that needs three retries.

On The X SubscriptionThe subscription Cursor is sold through: one plan for Grok, Cursor and X, with Free, Plus, Premium, Super and Ultra tiers, per-seat Teams, a weekly included pool for Cursor models shared with Grok, and third-party models billed at API list price from prepaid credit. Press Enter for the full definition., Grok 4.7Cursor and SpaceXAI's current lead Cursor model, in the weekly pool at $2 per million input tokens and $6 per million output, with a 256k context window and a 500k variant; above 256k input tokens it bills at 2x the standard rate. Press Enter for the full definition., Grok 4.6A Cursor and SpaceXAI model for complex coding and knowledge work, still in the Cursor model pool and succeeded by Grok 4.7; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition., Grok 4.5 and Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. draw from one pool that refreshes every week. The Grok apps and Grok BotAn early AI teammate product from SpaceXAI and Cursor that works across apps on a persistent cloud computer and asks for approval on selected actions; launch materials describe it as beta, but Cursor says Beta Services terms do not apply. Press Enter for the full definition. draw from the same pool, so Grok use outside the editor spends the budget you code with. A model from Anthropic, Google, OpenAI or Meta bills from prepaid credit at its listed API price, starting with the first request. Token rates differ by model, so two models can spend the same budget at very different speeds.

On Teams and Enterprise, the headline rate is not always the whole cost of third-party work. Eligible third-party requests add the Cursor Token Rate of $0.25 per million tokens. On Enterprise, a team that opts in to regional data residency also pays an uplift on eligible models. Keep the worked arithmetic in your rollout document, next to the model allow-list it depends on. Pricing explained covers the plan side of the sum.

On Teams and Enterprise, Auto runs on Cursor Router. Enterprise teams start with the router off, so an admin has to enable it first. Individual plans get only Cost mode, which is the classic Auto.

With the router on, you choose one of three modes (Cost, Balance or Intelligence), and the router selects the underlying model for each request. Cursor's docs say Cost keeps the bundled Auto pricing and bills as a Cursor-model request, though its Cursor Router help page says all Auto modes bill at the routed model's list price. Balance and Intelligence bill at the list price of whichever model the request goes to, so they can use up your usage faster than Cost.

• Use Cost when predictable bundled usage matters most.
• Use Balance when a team wants one practical default across mixed daily work.
• Use Intelligence for a task mix that justifies stronger models and a higher average spend.
• Pin a model when you need a reproducible comparison, a specific privacy approval or a known provider capability.

On Enterprise, the router respects the team's model access controls and routes only to allowed models. Cursor warns that blocking too many models reduces routing quality and can disable routing.

Decide the allow-list and the router mode together. When a team changes either one, rerun the same small evaluation set and compare cost, acceptance and review time again.

No. Privacy ModeCursor's setting that guarantees code data is not used for training by Cursor or its model providers, and that an admin can enforce org-wide; data-retention terms are a separate, contractual layer. Press Enter for the full definition. means neither Cursor nor its model providers train on your code, but retention and regional processing still vary by model and route. The Fable line is the documented exception. Cursor's pricing table says Claude Fable 5.1Anthropic's highest-priced Cursor model for long, complex work at $10 per million input tokens and $50 per million output, with $0.25 cache reads; Enterprise, Teams and Privacy Mode accounts need data-retention approval to use it. Press Enter for the full definition. needs data retention approval for Enterprise customers, and for Teams and individual customers with Privacy Mode enabled.

The approval exists because Anthropic stores agent input and output data for harm-prevention processes. That data is not used to train Anthropic models. Cursor's Fable 5.1 page says retained data is deleted after 30 days unless a safety investigation or legal hold applies, so Fable 5.1 does not fit a Zero Data Retention requirement.

Opus 5.5 sits on the other side of that line. Cursor's Opus 5.5 page says it is Zero Data Retention compatible, with no data-retention requirements for general access, so a team that standardized on Opus for ZDRZero Data Retention. A contractual guarantee that the model provider won't store your code or train on it. Press Enter for the full definition. can keep that setup. Don't infer a model's approval from its family name. Look up the exact model in Cursor's model pricing table, whose notes flag retention approval. Then check the region you selected and whether requests go through Cursor or your own API key.

Plus, at $8 a month, is the entry paid plan on The X SubscriptionThe subscription Cursor is sold through: one plan for Grok, Cursor and X, with Free, Plus, Premium, Super and Ultra tiers, per-seat Teams, a weekly included pool for Cursor models shared with Grok, and third-party models billed at API list price from prepaid credit. Press Enter for the full definition.. It runs Grok 4.7Cursor and SpaceXAI's current lead Cursor model, in the weekly pool at $2 per million input tokens and $6 per million output, with a 256k context window and a 500k variant; above 256k input tokens it bills at 2x the standard rate. Press Enter for the full definition., Grok 4.6A Cursor and SpaceXAI model for complex coding and knowledge work, still in the Cursor model pool and succeeded by Grok 4.7; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition., Grok 4.5 and Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. at fixed settings: Fast is off, effort is medium, and Grok 4.7 uses the 256k context window. Plus has no third-party models, Auto, Cloud AgentsAgents that run in a Cursor-managed virtual machine, check out the repo, do the work and open a pull request, then shut down, with no load on your laptop. Press Enter for the full definition., Automations or BugbotCursor's automated PR reviewer that posts inline findings and can push fix commits from isolated VMs. Press Enter for the full definition.. The choice of Fast mode and effort levels starts at Premium.

Cursor StartAn India-only legacy individual plan at ₹649 a month centered on Cursor's own models, with cloud and mobile access but without Other Models, Auto, Bugbot, Automations or the Agent SDK; Cursor names Plus as the comparable plan on The X Subscription. Press Enter for the full definition. is an India-only legacy plan, and Cursor names Plus as its comparable tier. Start includes Grok 4.7Cursor and SpaceXAI's current lead Cursor model, in the weekly pool at $2 per million input tokens and $6 per million output, with a 256k context window and a 500k variant; above 256k input tokens it bills at 2x the standard rate. Press Enter for the full definition., Grok 4.6A Cursor and SpaceXAI model for complex coding and knowledge work, still in the Cursor model pool and succeeded by Grok 4.7; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition., Grok 4.5 and Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. in non-fast mode, with the three Grok models at a fixed medium effort. It doesn't include Other ModelsOn legacy Cursor plans, the monthly pool for supported third-party models from providers such as Anthropic, Google and OpenAI, separate from Cursor Models. On The X Subscription third-party models are billed from prepaid credit instead. Press Enter for the full definition., Auto, on-demand usage, Automations, BugbotCursor's automated PR reviewer that posts inline findings and can push fix commits from isolated VMs. Press Enter for the full definition. or the Cursor SDK.

On either plan, use Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition. for ordinary work and switch to a Grok model when a job needs more depth.

Upgrade only when a missing capability has a real job attached to it, such as a third-party model your policy requires or Cursor Router for a team. Start adds one more reason: it has no on-demand usage, so it can't run past its included pool.

Test on your own work instead of a general coding benchmark. Choose three to five tasks your team already understands, such as a routine edit, a bug with a repro, a cross-file change and a task where the model should ask for clarification. Give each model the same starting context in a fresh conversation.

1. 1Write the acceptance check before the run: the test command, the expected behavior and the files that must not change.
2. 2Run the same prompt and starting context in a fresh conversation for each model.
3. 3Record elapsed time, tokens or cost, the test result, the files changed and the review minutes.
4. 4Reject any run that passes a test by weakening it, changes unrelated files or hides missing evidence behind a summary.
5. 5Choose a default by accepted changes and total review cost. Keep an exception for tasks where another model wins on the same check.

Rerun the set whenever a model, a price or the router policy changes, so the team's default rests on current evidence from its own work.