Which LLM Does Cursor Use in 2026? Models and Pricing

Short answer: Cursor does not use one LLM. It routes across first-party models (Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5) and, on paid tiers, third-party models from Anthropic, OpenAI and Google billed at API list price. Which one you actually get depends on your plan, your mode and whether you pinned a model or left it on Auto.

I went looking for this answer because a teammate asked why his Cursor bill changed shape overnight, and I realized I couldn't give a clean answer from memory. The docs shifted when Cursor became part of The X Subscription on October 12, 2026, and cached copies of the pricing page still disagree with each other. So I read the current pages and wrote down what is confirmed, what conflicts and what I couldn't verify.

One note on dates: I checked Cursor's pages in late October 2026. Cursor's docs don't show me an official last-updated date for the model list, so treat my timing as my own verification, not a vendor date. Model names and prices in this space move monthly.

What changed on October 12, 2026

Cursor announced that it is now part of The X Subscription. Current subscribers can keep their plan or switch. Per Cursor's announcement, third-party models move to API pricing on the new plans (Cursor on X). Existing subscribers keep their current plan with the same benefits and pricing, while new sign-ups and people upgrading from free Cursor land on the new plans (matt palmer on X).

This is why you will see two different Cursor price lists on the internet right now. One cached version of the docs still shows the older Pro $20 / Pro Plus $60 / Ultra $200 structure. The live version shows the new tiers. If a blog post, including a competitor's, quotes the old numbers, it is describing the legacy plans.

The model list, grouped by vendor

Cursor's docs say it supports frontier models from OpenAI, Anthropic, Google, SpaceXSI (spelled that way on the page) and others (Models & Pricing). Here is what the docs list, with context windows from the main docs page (Cursor Docs).

Two things I'd flag for anyone planning around context. First, "default" and "max" are different settings: the 1M ceiling exists, but you opt into it, and long-context runs cost more tokens. Second, Cursor's own models top out lower. If your workflow depends on dumping a monorepo slice into a single prompt, the third-party models are the ones with the 1M ceiling.

Pricing: first-party vs third-party

Here is the split that matters most. On the new plans, Claude, GPT and Gemini are not part of the included usage pool. On Premium and above they are billed at the model's API list price from the first request, paid from prepaid credit (Models & Pricing).

Individual plans:

Teams seats cost $30, $100 or $200 per seat per month, matching the Premium, Super and Ultra tiers. Third-party models on Teams are billed at API list price plus the Cursor Token Rate on eligible models (Models & Pricing). The docs mention the rate, but I did not find an official figure I could confirm, so I won't quote one. Check the pricing page for the current rate.

Per-million-token list prices that the docs show for third-party models:

The price tables in the search output lost their column headers. I read the numeric columns as input and output per 1M tokens, which is the standard layout, but check the live page before you build a budget on this. The docs table also has more columns (I believe cache write and cache read) and an unlabeled Yes/No column whose meaning I can't confirm.

For first-party models, Grok 4.7 costs $2 input and $6 output per 1M tokens, and Grok 4.7 (Fast) costs $4 and $12. Composer 2.5 costs $0.50 input and $2.50 output per 1M, while its faster variant, which has the same intelligence, costs $3.00 and $15.00 and is the default option (Composer 2.5 announcement).

What Free and Plus actually get you

The Free and Plus tiers include only Cursor's first-party models: Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5. Premium and above, Teams, Enterprise and the legacy Pro, Pro+ and Ultra plans unlock third-party models (Available models).

Cursor's help center calls Grok 4.7 its flagship, built for the hardest and longest-running tasks. On Plus, the first-party models run at fixed settings: Fast is off, effort is medium, and Grok 4.7 uses the 256k context window (Models & Pricing).

So on the $8 plan you are not choosing between Claude and GPT. You are choosing among Cursor's own models, with the knobs locked. That is fine for a lot of day-to-day editing, and I'd rather someone know it up front than discover it when they try to select Opus and can't.

Composer 2.5 deserves a sentence of its own. It was released May 18, 2026, about two months after Composer 2 (March 19, 2026), and Cursor's blog says it is built on Moonshot's open-source Kimi K2.5 checkpoint (DataCamp's write-up). At $0.50/$2.50 per million tokens in its standard variant, it is by far the cheapest model in this lineup. If you run agentic loops that burn tokens, that gap matters more than any benchmark.

Auto and the router: where the answer gets fuzzy

If you leave the picker on Auto, you are not selecting a model; you are letting Cursor Router pick. The help center lists the models the router routes across as GPT-5.5, Claude Opus 5, Grok 4.6 and Claude Fable 5, and says Grok 4.6 is required: blocking it disables the router (Available models).

Cursor's routing guide says Auto is backed by Cursor Router with Cost, Balance and Intelligence modes, and Balance and Intelligence bill at the routed model's list price (Model routing guide).

Here is what I could not confirm, and I'd rather say so than guess:

• Which model is the "required" one on older pages. The help page says Grok 4.6. I wasn't able to confirm which model older Cursor pages name as required, so check the current help page rather than relying on cached copies.
• Whether the router applies to individual plans. The guide describes it as being on Teams and Enterprise. The help page describes Auto Cost as billing from your included pool. I couldn't confirm how Auto behaves on individual X Subscription tiers.
• The default mode for new users. I didn't see a default mode stated in the official pages, so I'm not stating one.
• A default model per plan. I found no official page that says "the default on Free is X, on Premium is Y." Anyone who gives you that table with confidence is guessing or reading a cache.

That last point is why the "defaults per tier" table you may have seen elsewhere isn't in this post. What is confirmed: Free and Plus are limited to first-party models, Auto is the router entry point, and Grok 4.7 is the flagship.

What I'd do

I've spent years building systems that call LLM APIs unattended, and the habit that transfers to Cursor is simple: treat the model as a cost line, not a vibe.

1. Start on Plus if you're evaluating. At $8/month you get Grok 4.7 and Composer 2.5, which is enough to learn whether agentic editing fits your workflow. Don't pay for third-party access until you can name a task the first-party models fail at.
2. Move to Premium only when you can name that task. Third-party models bill at API list price from the first request out of prepaid credit. That is pay-as-you-go economics, so a single long-context Opus run costs real money. Set a mental per-task budget.
3. Pin a model for expensive work, use the cheap one for loops. Composer 2.5 at $0.50/$2.50 per million is the right default for repetitive agent iterations. Reserve the $10/$50 tier for the one hard problem per week.
4. Don't trust any pricing table that lacks a date. Including mine. Open the live pages (Models & Pricing, Available models) before you commit a team budget.
5. If you're on a legacy plan, check before switching. Cursor says you can keep it with the same benefits and pricing. Whether the new structure beats it depends on how much of your usage was third-party models.

The pattern underneath all of this is the same one I use when I build agent pipelines: route cheap models to high-volume steps, reserve expensive models for the verification and hard-reasoning steps, and measure what each step actually costs. Cursor's new pricing makes that discipline mandatory instead of optional.

If you're sizing an AI tooling budget for an engineering team, or wiring models into your own product and want a second pair of eyes on the cost architecture, you can reach me at lazar-milicevic.com/#contact. Otherwise, there's more production-focused LLM writing on the blog.

I checked Cursor's live pages in late October 2026. This is my own verification timing, not a vendor-published update date.