Review · 2026-09-17

Cursor review

4.5/ 5

Written by a competitor: we build TODO for AI. What that means for this review.

The best AI editor available, and still the fastest way to write code with a model at your side. Its limits are the limits of a repo-shaped tool, not of an editor that cannot leave your screen.

Rated as an AI code editor, the job it was actually built for. Reviewed against Pro, Pro+ and Ultra plans as priced in September 2026.

4.5
our rating
$20/mo
from
IDE
interface
2026-09-17
reviewed
Disclosure

Who is writing this

We are not neutral
We build TODO for AI, which competes with Cursor for some of the same budget. We pay for Cursor and use it. This review says where it beats us, and we have kept our own case to one clearly marked note at the end.
Method

How we reviewed it

  • —Hands-on use of Cursor on a large TypeScript and Python monorepo through August 2026, alongside the CLI agents we build. We have not used it daily since, so anything shipped in the last few weeks is not covered here.
  • —Pricing, Cloud Agents and browser behaviour read from Cursor's own docs on 17 September 2026, not from third-party summaries.
  • —We have not run a controlled benchmark against other editors. Where we say something is fast or accurate, that is a working impression, not a measurement.
The case for

What Cursor does well

Tab completion is still the best in the category

It predicts the edit you were about to make, across lines and across files, rather than the token you were about to type. It is our favourite feature in any AI tool and the single thing most likely to keep you subscribed. Cursor includes it unlimited on every paid plan.

A familiar VS Code foundation

It is a VS Code fork with import for your extensions, keybindings and settings, so the cost of trying it is close to zero. It lags upstream VS Code by a version or two and some UI differs, but that low switching cost is a large part of why it won the category.

Models from several vendors, one subscription

Claude, GPT, Gemini and Cursor's own Composer, switchable per request. You are not locked to one lab's idea of a good model, which matters more each time the frontier moves.

Cloud Agents genuinely run without you

Cursor runs agents in its own isolated VMs that, in its words, "do not require your local machine to be connected to the internet". They clone the repo, work on a branch and open a merge-ready PR. Anyone still saying Cursor stops when you close the laptop has not looked recently.

The case against

What frustrates

The pricing change was handled badly, and the shape of it still confuses people

In June 2025 Cursor replaced 500 monthly requests with a pool of API-priced credit, and described it in a way that led people to believe far more usage was included than was. They apologised publicly, offered refunds for unexpected charges in the affected window, and rewrote the docs. Credit where it is due: that is a better response than most vendors manage. The two-pool model it left behind still takes a careful read.

Source: Cursor, "Clarifying our pricing", July 2025
Only Tab is unlimited, and Auto is not the escape hatch people think

Unlimited tab completions come with every paid plan. Agent usage does not: included usage is split into a Cursor-models pool and an other-models pool, and Cursor is explicit that "all Auto modes bill at the list price of the model each request is routed to". Auto picks the model for you, it does not stop the meter. Your model choice, not your seat, decides how fast the month burns.

Source: Cursor pricing docs, September 2026
Long agent runs drifted for us on a big codebase

Our experience through August 2026, not a measurement: on our monorepo, foreground agent sessions working through a multi-file change tended to lose the thread of earlier decisions and produce edits that did not match the conventions around them. Review caught it and it was still faster than doing it by hand, but we have not tested whether this happens more in Cursor than in other coding agents, and we have not put the same work through Cloud Agents at length.

Everything is shaped like a repository

Cloud Agents start from a connected repo and finish at a pull request. The browser tool exists so the agent can check the app it just changed — navigate, click, screenshot, read the console. Both are good. Both assume the unit of work is a code change. If the task is answering an inbox, updating a CRM or running an ad account, there is no branch to open and no PR to merge, and that is a deliberate scope decision rather than a gap.

Source: Cursor Cloud Agent docs, September 2026
Money

Pricing, honestly

Pro $20, Pro+ $60, Ultra $200 per month. Teams starts at $40 per user per month, with Teams Premium at $120 and Enterprise custom.

Unlimited tab completions on every paid plan. Agent usage is not unlimited: it draws on an included Cursor-models pool and an other-models pool that reset with your billing cycle, and Auto bills at the list price of whichever model it routes to.

Third-party models bill at the provider's API rates rather than a marked-up per-request price, which is more honest than it is usually given credit for. On Teams and Enterprise a Cursor Token Rate of $0.25 per million tokens applies on top, including on BYOK usage.

The unpredictability is inherent to token pricing, not a margin grab. But it does mean the invoice depends on which model you reached for, so the plan you need is a function of your habits rather than your headcount.

The answer

Who should buy it

Buy Cursor if
  • You write code most days and want the best assistance available while you type
  • You live in VS Code and want to keep your setup
  • You want to switch models per task without switching tools
  • Your delegated work is repository work that ends in a pull request
Look elsewhere if
  • Most of your week is not source code
  • You want one predictable number on the invoice regardless of which model ran
  • You need work that ends somewhere other than a branch: an inbox, a CRM, an ad account
Where we fit
If you mostly type, Cursor is the right purchase and we would not try to talk you out of it — we pay for it too. The honest split is not presence versus delegation, because Cursor delegates fine when the task is a pull request. It is the shape of the work: TODO for AI is built for tasks that do not end in a branch, running on your own machine or a cloud VM across a browser and business tools. If your backlog is all code, you may not need us at all.
Full feature comparison vs TODO for AICursor alternatives
FAQ

Common questions

Is Cursor worth it in 2026?
If you write code most days, yes. Unlimited tab completion alone justifies $20 a month. Budget separately for agent usage, which is metered by model rather than included outright.
Is Cursor's Auto mode unlimited?
No. Unlimited tab completions come with every paid plan, but Cursor's docs state that all Auto modes bill at the list price of the model each request is routed to. Auto chooses the model for you; it does not stop the meter.
Why did people get unexpected Cursor bills?
A June 2025 change swapped 500 monthly requests for a pool of API-priced credit, and the announcement implied more was included than was. Cursor apologised in July 2025 and offered refunds for unexpected charges in the affected window.
Can Cursor work while my laptop is closed?
Yes. Cloud Agents run in Cursor-managed VMs and do not need your machine connected. They clone a repo, work on a branch and open a pull request, so the work has to be shaped like a code change.
Is Cursor good for large codebases?
For editing with you present, yes. For long multi-file agent runs, our experience through August 2026 was that edits drifted from surrounding conventions and needed review. We have not benchmarked that against other agents.