What AI model does it use?

What AI model does Cursor use?

A year ago most people would have answered "Claude, mostly". Cursor's own models page, read on September 25, 2026, puts something else first. It splits everything into two usage pools, and the one with the bigger included allowance is called Cursor Models: Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5.

Why Grok is a Cursor model now

Cursor was bought by SpaceX. The company's own post of August 14, 2026 says "Cursor has officially been acquired by SpaceX", and since then the Grok releases have come out of Cursor as well as xAI's side of the company (its pages now call it SpaceXAI). The newest, Grok 4.7, shipped on September 21, 2026, and Cursor describes it as "a frontier model from Cursor and SpaceXAI for coding and knowledge work."

That ownership shows up in the bill. The models page says first-party Cursor models, "including Grok and Composer, are exempt from the Cursor Token Rate", which is the extra per-token charge Teams and Enterprise pay on third-party models.

Composer, the in-house model, and what it's built on

Composer is the model Cursor trains itself for agent work (an agent here meaning the mode that edits files and runs commands across several steps). Composer 2.5 came out on May 18, 2026.

What surprised me is in the launch post itself: "Composer 2.5 is built on the same open-source checkpoint as Composer 2, Moonshot's Kimi K2.5." So Cursor's own model starts from another lab's open-weights model, which Cursor then post-trains (further training on its own coding tasks). Cursor says it's training a much larger model from scratch with SpaceXAI, but that one isn't out.

The third-party models you can pick

The second pool is Other Models, billed at each provider's API price. On September 25 it held Claude Sonnet 5, Claude Opus 5.5 and Claude Fable 5.1 from Anthropic, GPT-5.6 in its Sol, Terra and Luna sizes from OpenAI, Gemini 3.1 Pro and Gemini 3.8 Flash from Google, and Meta's Muse Spark 1.3. The pricing table also carries Z.ai's GLM 5.2 and Moonshot's Kimi models.

So yes, Cursor still uses Claude. It just isn't the house model any more, and you only get it by choosing it (or when Auto routes to it).

What Auto picks

Auto is the setting where Cursor chooses for you. It has three modes, Cost, Balance and Intelligence, and on Teams and Enterprise the last two run through Cursor Router, launched July 22, 2026. The router docs are frank that you don't get to see inside it: "You can't hand-pick which model handles a request, and the model pool changes over time as new models ship."

By default the router doesn't even show which model answered. "Hidden is the default and recommended, so results are judged on their own merit rather than by model name," the docs say, and an admin can switch the display on. One model is required, though: "enabling Grok 4.6 is a requirement for the router to work", because it's the cheap workhorse the router leans on.

What depends on your plan

Pro, Pro Plus and Ultra get both pools. Start, Cursor's lower-priced plan for developers in India, gets only the Cursor Models pool, and the docs spell out that it doesn't include "the Other Models pool, on-demand usage, Bugbot, Auto, Automations, or the Cursor SDK." On Start, then, the honest answer is Grok or Composer and nothing else.

What Cursor doesn't say

Cursor names no single default model. It also doesn't publish which models Auto's Cost mode chooses between on individual plans (the docs say only that it "uses the previous Auto routing logic"), and the models page doesn't mention the free Hobby tier. If you want to know exactly which model wrote a piece of code, pick one by hand.