Cursor Models: The Two Pools and How to Pick One

The two pools

Cursor Models

Included generously with paid plans. This is the pool Auto draws from, and the one your monthly allowance is designed around.

Grok 4.7 xAI: Default frontier model in the Cursor pool. A 500k-context variant and a Fast variant are billed at higher rates.
Grok 4.6 xAI: Previous Grok generation, kept for workloads already tuned against it.
Grok 4.5 xAI: Oldest Grok still in the pool.
Composer 2.5 Cursor: Cursor's own model. The cheapest option in the pool and the one Hobby accounts get.

Other Models

Third-party frontier models, billed from a separate pool at rates closer to what the providers charge directly. Powerful and considerably easier to exhaust.

Claude Opus 5.5 Anthropic: Tops CursorBench 4.0 at every reasoning effort. The most expensive way to run a long task.
Claude Fable 5.1 Anthropic: Anthropic's mid-weight line, second on CursorBench behind Opus 5.5.
Claude Sonnet 5 Anthropic: The lighter Anthropic tier, still listed alongside Opus and Fable.
GPT-5.6 Sol, Luna and Terra OpenAI: Three variants of the same generation, each with selectable reasoning effort.
Gemini 3.1 Pro Google: Google's frontier entry, selectable from the picker and the CLI.
Gemini 3.8 Flash Google: The fast, cheap end of Google's line-up.
Muse Spark 1.3 Meta: Meta's entry in the list. The docs page hides further models behind a Show more control.

Why the split matters

Most confusion about Cursor usage comes from treating the allowance as one number. It is two. Someone who leaves the picker on Auto all month may never get close to a limit, while someone who pinned a frontier model on day one can run dry in a fortnight on the same plan. Neither is a billing error.

The practical consequence is that your model habit matters more than your tier. Upgrading from Pro to Pro+ triples what you can spend; switching from a pinned frontier model to Auto can achieve a similar effect for nothing.

What Auto is actually doing

Auto is not a compromise setting. It routes each request to a model in the Cursor Models pool based on what the request appears to need, which means the cheap model handles the many turns that are really just edits, renames and file lookups, and something stronger handles the turns that are not.

The case for overriding it is narrow and specific: a genuinely ambiguous refactor across several files, a bug that has already resisted two attempts, or a design decision where you want the best reasoning available regardless of cost. Outside those cases, pinning a frontier model mostly buys you a faster route to your usage ceiling.

Auto is not a single behaviour either. It has three optimisation modes, handled by what Cursor calls the Router, and on Teams and Enterprise plans the Router chooses the model for every Auto request according to the mode you set.

Cost. Routes towards the cheapest model that can plausibly handle the request.
Balance. The middle setting, trading spend against capability per request.
Intelligence. Routes towards capability, accepting a higher bill per request.

Whichever mode is active, each request bills at the list price of the model it was routed to. Auto is a spending strategy, not a flat rate.

The surcharge teams pay

One line in Cursor's pricing documentation catches teams out. On Teams and Enterprise plans, third-party model requests carry a Cursor Token Rate of $0.25 per million tokens, applied on top of the model's own API pricing. It applies to included usage, on-demand usage, and to requests that Auto routed to a third-party model without you asking it to.

First-party models are exempt. Grok and Composer do not carry the surcharge, which quietly widens the cost gap between the two pools for any organisation large enough to notice a quarter of a dollar per million tokens.

What the benchmark actually says

Cursor publishes CursorBench, scored on ambiguous multi-file tasks drawn from real sessions rather than synthetic puzzles. Three rows from the top of it are worth reading together, because the pattern they form is more useful than the ranking.

Opus 5.5 Max scores 57.8 percent at $13.43 per task.
Opus 5.5 High scores 56 percent at $3.97.
Opus 5.5 Low scores 43.7 percent at $1.17.

The gap between the first two is 1.8 percentage points for roughly 3 times the money. The third costs about a tenth of the first and still lands within 14.1 points of it. Reasoning effort has sharply diminishing returns, and the top of any leaderboard is rarely the rational default.

Prices are deliberately not mirrored here

Per-token rates change frequently enough that any static copy is a liability. The official models and pricing table is the only version worth quoting, and that applies to this page as much as to anyone else's.

Model line-up and benchmark figures read from CursorBench and the models documentation on 25 September 2026.

Frequently asked questions

What is the difference between Cursor Models and Other Models?

They are two separate billing pools. Cursor Models covers the models Cursor includes generously with paid plans, and Other Models covers third-party frontier models billed closer to API rates. Exhausting one does not exhaust the other, which is why a plan can feel simultaneously generous and restrictive depending on what you have pinned.

What model does Cursor Auto pick?

Auto routes each request to a model from the Cursor Models pool based on the request, favouring cheaper models for work that does not need frontier reasoning. It is the right default because most agent turns are edits and lookups rather than hard reasoning problems, and it is how a limited allowance stretches furthest.

Which Cursor model is best for coding?

On Cursor's own CursorBench 4.0 leaderboard, Opus 5.5 Max leads at 57.8 percent, at an average 13.43 US dollars per task. The more useful comparison is Opus 5.5 High, which scores 56 percent for 3.97 dollars: roughly the same result for about a third of the cost.

Is Composer worth using?

For a large share of everyday work, yes. Composer is Cursor's own model and the cheapest thing in the pool, and it is what Hobby accounts get. Its weakness is genuinely ambiguous multi-file reasoning, which is exactly the case where switching to a frontier model pays for itself.

Does a higher reasoning effort always produce better results?

No, and the leaderboard shows it. Moving Opus 5.5 from High to Extra High changes the benchmark score by nothing measurable while roughly doubling the cost per task. Effort levels buy more thinking, not more correctness, and the returns flatten well before the top setting.

Why is a model missing from my model picker?

Three common reasons: your plan does not include that pool, your editor build predates the model's launch, or your team's administrator has restricted the available models. Updating the editor is the fastest thing to rule out, and `agent models` from the CLI lists exactly what your account can call.

How much does each model cost per token?

Cursor publishes a per-token table that changes often enough that any copy of it goes stale quickly. Read it on cursor.com/docs/models-and-pricing rather than trusting a third-party mirror, including this one.

What are the three Auto modes?

Cost, Balance and Intelligence, handled by Cursor Router. On Teams and Enterprise plans the Router picks the model for each Auto request according to the mode you have set. Every mode bills at the list price of whichever model it routed to, so Auto changes your average spend rather than fixing it.

What is the Cursor Token Rate?

A surcharge of $0.25 per million tokens applied to third-party model usage on Teams and Enterprise plans, on top of the model's own API price. It covers included usage, on-demand usage, and Auto requests that get routed to a third-party model. Cursor's own models, meaning Grok and Composer, are exempt.

What is Max Mode?

A legacy feature. Max Mode extends a model's context window beyond its default limit and bills at the model's API rate plus 20 percent, and it is available only on legacy request-based plans. Current plans do not have it.