Which AI Models Does Cursor Support? Composer, Fusion, and Frontier Models

One of the first decisions you make in Cursor is not what to build, but which model to build it with. Open the model picker and you are looking at a menu that spans four outside providers plus Cursor's own engines. That flexibility is a real selling point, and it is also a small source of confusion: when several models can all do the job, how do you pick the right one for the task in front of you?

This breakdown lays out exactly which models Cursor supports, grouped by who makes them, and then gives you a practical way to choose. The list below is current as of June 9, 2026 and it moves fast, so treat it as a snapshot and confirm the live roster in Cursor's model documentation. If you are new to Cursor itself, start with what Cursor is first, then come back here.

How Model Choice Works in Cursor

Cursor is model-flexible by design. The product framing is blunt: choose between every leading model from OpenAI, Anthropic, Gemini, and xAI, or Cursor's own. In practice that means you bring the model to the task rather than being locked to a single provider, and you can switch models without leaving the editor.

Two parts of Cursor have a model behind them. The Agent, which reads your codebase, edits files, and runs commands, can run on Cursor's in-house Composer for low latency, or on any of the third-party frontier models when you want their particular strengths. Cursor Tab, the fast autocomplete that predicts your next move, is powered by Cursor's own Fusion model. So the choice you make in the picker mostly governs the Agent; Tab quietly runs on Fusion underneath.

Because Cursor connects to outside providers, model availability and usage cost are tied to those providers and to your plan's usage-based billing. That is part of why the roster changes often, and why the honest answer to "which model should I use?" is "it depends on the job," which is what the rest of this breakdown is about.

The Frontier Models You Can Bring

The third-party side of the picker pulls in flagship models from four providers. These are the general-purpose, deep-reasoning models you reach for when a task needs strong planning, tricky refactors, or careful problem-solving rather than raw speed.

A few of these names hint at their intended use. OpenAI's GPT-5.3 Codex and xAI's Grok Build 0.1 are coding-leaning variants from their families, while the standard releases like GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.8 are broad reasoning models. Cursor does not lock you into one, so the right move is to try a couple on a representative task and keep the one that handles your codebase best. We are deliberately not attaching benchmark scores or per-model prices here, because those are not in our verified sources and they shift quickly.

Cursor's Own Models: Composer and Fusion

The reason the in-house models exist is speed and integration. Third-party frontier models do strong work but route over the network to another company; Cursor's own models are tuned to feel instant inside the editor and to understand your whole project through codebase-wide semantic search.

Composer

Composer is Cursor's low-latency agentic coding model, built to be deeply integrated with the editor and trained with codebase-wide semantic search. It is what makes the Agent feel quick on routine work. The line has iterated fast: Composer arrived in October 2025, then Composer 1.5 in February 2026, then Composer 2 in March 2026 (built on a Kimi K2.5 base), and the current Composer 2.5 in May 2026.

Fusion

Fusion is the in-house model behind Cursor Tab. Announced in January 2025, it powers the fast autocomplete that finishes your current line and predicts your next action, including the "cursor jumps" that anticipate where you will move next in the file. You do not pick Fusion from the model menu the way you pick a frontier model; it runs under Tab automatically, which is why Tab feels instant even when a heavier frontier model is handling the Agent.

The Full Model List, by Provider

Here is the complete list grouped by provider, as of June 9, 2026. Because this changes frequently, the live source of truth is Cursor's documentation rather than any article, including this one.

Max Mode and Long-Context Work

Some tasks need the model to hold a lot of your codebase in mind at once: tracing a bug across many files, reasoning about a large refactor, or summarizing a sprawling module. For those, many models in Cursor support a Max Mode that extends the context window up to 1M tokens.

Two caveats keep this honest. First, Max Mode is supported by many models but not necessarily every one, and which models qualify can change, so check the model documentation rather than assuming. Second, a bigger context window is a tool, not a default: feeding the model more than the task needs can slow things down and add usage cost without improving the answer. Reach for Max Mode when the problem genuinely spans a large body of code, and keep it off for the everyday edits where a tighter context is faster and cheaper.

Which Model for Which Job

Strip away the model names and the decision comes down to three jobs. Match the job to the kind of model rather than chasing whichever release is newest.

If you want the mechanics of how the Agent uses whichever model you pick, see the Cursor Agent; for how the Fusion-driven autocomplete behaves, see Cursor Tab. For how model usage maps to what you pay, see Cursor pricing.

Honest Limitations

The model picker is a strength, but it comes with a few realities worth naming before you build a workflow around any one model.

Frequently Asked Questions