When the update never applies
The most reported update problem is not a failed download. It is an update that downloads successfully, prompts for a restart, and then reappears afterwards asking to be installed again. The cause is nearly always that the application cannot overwrite itself.
On macOS, this happens when Cursor is running from somewhere other than /Applications: a Downloads folder, a Desktop, a second disk. The updater writes the new build, fails to swap it in, and silently falls back. Moving the application into /Applications and launching it from there usually fixes the loop permanently.
On Windows, the usual cause is having both the System and User installers on the same machine. Two installations means one updater can be patching a copy you are not actually running. Uninstall both, then install one deliberately.
In either case, installing the current build over the top from the official download page resolves it without losing anything. Settings, keybindings and extensions live in your user profile, not inside the application, so a reinstall does not touch them.
Holding a version on purpose
Occasionally a release changes behaviour you depend on, or breaks an extension. Staying put is legitimate, and needs two steps rather than one: install the specific build you want from the version list on the download page, and turn automatic updates off. Doing only the first means the application quietly restores the version you were trying to avoid, usually overnight.
On Linux with .deb or .rpm packages, this is a package manager question rather than an application setting. Hold the package the way you would hold any other, and the editor will not argue.
The CLI updates on its own terms
The agent CLI tries to auto-update so you stay current. That is the right default on a laptop and the wrong one in a container, where a version that changes between runs turns a reproducible pipeline into an unreproducible one. In images, install at build time and treat the runtime updater as something to work around rather than rely on.
A missing model is often a stale client
Model availability depends on the client as well as the account. A build that is a few releases behind can be missing a model that launched recently, and the symptom looks like a plan problem rather than a version problem. If the picker does not show something you expect it to, update before you open a support ticket.
Behaviour described against the download page and the CLI installation docs as of 25 September 2026.