Cursor is dead to me
I opened Cursor this morning out of habit. The window sat there, idle, while I reached for my terminal instead.
A few years ago, that would've been unthinkable. Cursor was the editor -- their tab-editing was the best in the game, bar none. Speedy, smart, and fully context-aware. It understood what I was about to type before I even knew I was about to type it. For a long stretch, it was my single biggest productivity advantage. No contest.
But it's 2026 now, and something fundamental has shifted. I'd argue that mechanically coding -- actually typing code character by character -- is no longer the most productive thing I can do. Frontier models can code faster and better than I can, especially when the task is well-bounded and the constraints are clear and unambiguous. The gap isn't small anymore. It's a chasm.
In the spirit of me going a full 360 on my agentic anthropomorphization -- I've gone from insisting we shouldn't pretend LLMs are software engineers to admitting I'd be full of boot for that stance -- this evolution feels like the next logical step. First I stopped pretending. Then I stopped coding entirely.
And that shift killed Cursor's core appeal for me. If hand-coding has become rote -- something I delegate, not something I do -- then a powerful tab-completion tool starts to feel like a beautifully engineered horse-drawn carriage in the age of cars. Sure, it's the best carriage. But I'm not trying to ride a horse anymore.
I'm aware of the irony, by the way. I once wrote that dumb Cursor is the best Cursor -- and I still think that take was right for its time. Stripping Cursor down to a surgical edit-only tool was the best way to make it useful back then. But, oh well, here we are now.
What Cursor still gets right
I don't want to bury Cursor entirely. It's not a bad product -- it's actually still quite good.
The Composer family delivers real value at breakneck speeds, and its coding quality remains impressive. The agents workflow is solid, even if I don't personally reach for it much. And their model selection is genuinely great -- you get (allegedly) 2x the value of your subscription on their hosted models.
But a good product isn't the same as an essential one. And somewhere along the way, Cursor stopped being essential for me.
And here's one small thing that drives me absolutely nuts: Cursor CLI keeps injecting Made-with: Cursor on every commit and PR description. Every. Single. Time. Even when I explicitly tell it not to. It's a tiny thing, but it's the kind of friction that accumulates when you're already questioning whether a tool still belongs in your workflow.
The contenders showed up
What pushed Cursor from "daily driver" to "occasionally opened" was the rise of tools that don't ask me to code at all.
OpenCode is my current daily driver, and honestly, it's been incredible. The model selection through its Go and Zen tiers is excellent -- I get access to the models I actually want without having to think about it. But the real killer feature is the multi-frontend experience: I can use it over CLI, over SSH, and through a surprisingly capable web UI. Since my dev workstation is on Tailscale, I can ship changes from literally anywhere -- heck, even from my phone. The web UI actually works on mobile, which is more than I can say for Cursor's web agents dashboard, where the text entry is completely hidden on a tablet. Same session, same context, same quality.
Man, I didn't realize how much I needed that until I had it.
The nail in the coffin
And then OpenAI drove the last one in.
I subscribed to ChatGPT Pro at $100/month, which -- for my use case -- gives me virtually unlimited usage. gpt-5.5 handles the tough, complex refactors. gpt-5.3 codex spark chews through the smaller, more trivial tweaks. Both are bundled in.
But here's the part that really sealed it: OpenAI lets you use the Codex CLI as a provider backend in OpenCode. I can sit in my OpenCode workflow, fire off a task, and it runs against my ChatGPT Pro subscription. No separate billing. No juggling credits. No wondering if I'm about to hit a rate limit mid-session.
At that point, what am I keeping Cursor around for?
Claude, sidelined
Claude remains a strong competitor on raw capability. I still think Claude Code is arguably the best at truly complex, multi-phase work -- the kind where you need a plan, an implementation, and an adversarial review before anything gets committed.
But Anthropic is Scrooge McDuck. They don't let you use the Claude CLI as a provider in other tools -- or rather, they technically do, but then they bill you for extra usage anyway, which makes the whole thing feel cosmetic at best. So while I still fire up Claude Code for the heavy stuff -- usually in a ping-pong workflow where Claude and Codex trade plans and reviews until they converge -- it's taking more of a backseat in my day-to-day.
It's a shame. Anthropic makes great models. They're just weird about letting you use them where you actually want to.
So yeah. Cursor is dead to me. Not because it got worse -- the product is still solid. But because my relationship with code changed, and Cursor didn't change with it.
Or maybe it couldn't. A tool built around making you a faster typist doesn't have much to offer when you stop typing.