Cursor's Composer Model Surprised Me

I started using Cursor agent again for the first time since March, and Composer 1 honestly surprised me.

I now split most coding tasks between GPT-5 Codex and Composer. I could actually see Cursor becoming a daily driver again, not because it's the "smartest," but because it's so fast and keeps you tightly in the loop.

The Speed vs. Intelligence Tradeoff

For a long time I assumed the smartest possible model was always the right tool. But it turns out that isn't always true. I'm still figuring out the balance between using a slightly dumber model that is extremely fast versus a smarter model that adds latency. Ideally there will be a future where the smartest models are also instant, and this tradeoff disappears entirely. We are not there yet, but Composer 1 feels like a real step in that direction.

There's a whole class of work where you already know exactly what needs to change: surgical edits across a few files, quick refactors, small rewrites. Composer 1 absolutely shines here. It's fast enough that staying human in the loop stops feeling like a tax and starts feeling like leverage.

Intelligence-wise, it's roughly in the Haiku-ish tier (maybe a bit above Anthropic's latest Haiku 4.5), but for these targeted workflows the speed matters more than the raw IQ. Honestly, I don't reach for Anthropic at all anymore. It's basically GPT-5 Codex for deep or long tasks and Composer when I want to stay tight to the code.

Days of Future Past

Using Cursor again feels like living in the future and the past at the same time. The code review flow is still second to none. But parts of the UI feel like relics from 2023 and 2024, back when you had to manually approve everything because models weren't consistently good.

After spending months YOLO-ing long tasks with Claude Code and Codex, the "approve this, approve that, review every diff" workflow feels both comforting and a little antiquated. Models are valid more often now, so the old guardrails start to look like overhead.

Still, there's something valuable there. It feels like Cursor is at a moment where they will eventually have to kill some of their babies, trimming features that were essential in the early days but now hold back the speed they are clearly optimizing for. With a user base their size, I get why it is hard.