Choosing a LaTeX Editor in 2026: The Questions That Decide It

Feature comparison tables are written by whoever benefits from them, and they go stale in a quarter. Rather than another one, here are the five questions whose answers actually change which editor you should be using, and how to check each yourself in a few minutes. Where the answer is ours, it is stated plainly. Where it is somebody else's, go and look, because it may have moved since this was written on 2 September 2026.

1. Do collaborators cost money?

This is the question that decides it for most groups, and it is rarely at the top of a comparison table. Per-seat pricing means a six-author paper is six subscriptions, and the person who pays is usually the one who least wants to administer it. On LetX collaborators are unlimited and free: the paid tier is about AI usage, not about how many people may open the document. If you are evaluating anything else, find the number of people who can edit one project simultaneously on the free plan, because that is the cap that bites first. Our breakdown of the Overleaf free plan goes through those limits in detail.

2. Is your unpublished work training data?

For a paper under review this is not an abstract concern. Check two things: the terms, and the site's own robots.txt, where Content Signals now let a site state its position in a machine-readable way. letx.app declares ai-train=no and is not tied to any single model vendor. A service that says nothing has not said no.

3. What happens when a compile gets long?

Every hosted editor has a limit somewhere, and a 200-page thesis with a bibliography and forty figures is where you meet it. The honest number for us, measured end to end rather than for the compile step alone, is a few seconds for an ordinary document: a queue median around two seconds and a compile median around three and a half. Any figure much smaller than that is measuring a single stage and presenting it as the whole trip, so when you see one, ask what it timed. What you want to know about any editor is what happens when a build exceeds its ceiling, and whether you can still get a PDF out.

4. Is version history usable on your actual plan?

History is the feature everybody assumes and nobody tests until they need it. The questions worth asking are how far back it goes, whether it is available without paying, and whether you can see a diff rather than only restore. Test it by making an edit, waiting, making another, and trying to recover the middle state. Do that before you need it, not during.

5. Can you leave?

LaTeX is plain text and some images, so there is no technical reason any editor should be hard to leave. That makes it a good integrity test: download a zip of a real project and compile it somewhere else. If that works, you are choosing on merit rather than on lock-in, which is the position you want to be in every time the pricing changes.

The things that are new this year

Two capabilities did not exist in this category eighteen months ago and are worth weighing separately from the five questions above, because they change the daily work rather than the terms of the deal.

An assistant that edits the document rather than describing it. The distinction is whether it can read your project, propose a diff you accept or reject, and then compile to check its own work. An assistant that cannot compile is guessing, and what you ask it decides which one you get.

Reaching the editor from the assistant you already use. If you live in ChatGPT or Claude, a remote MCP server means the compiler comes to you: one URL and no install.

If you are deciding for a department

Different question, different answer. Per-person verification does not scale to a faculty, and neither does per-seat billing. See choosing an editor for a university and the academic pricing paths, which cover site licences and why a .edu pattern match excludes most of the world.