My introduction to LLM-assisted coding was using Cursor, towards the start of 2025 (might have been late 2024).
At the time, it was still a relatively new fork of VS Code, but rapidly grew in functionality and stature within the developer community.
The Cursor product has made great strides since then, and I’ve stuck with it through a number of UI/UX misteps and other issues. I never bought into the “agent only” model of working they were heading towards—I still need to be able to access an editor for the work I do, and how I operate.
But I have appreciated the relative simplicity of the workflow and advancements in their own model (Composer) and the “Auto” mode that removed a lot of the manual consideration required in choosing the right model. The code caching mechanism means it’s often faster for me to use an LLM to find something in the large codebase I work with in my day job.
I have basically settled on a Composer 2.5/Claude Opus 4.8 model combo, and I’ve become quite productive with that pairing, in the Cursor environment.
However, about a month ago murmurings started to surface about an acquisition by SpaceX, a company that is significantly owned by Elon Musk. This has since come to pass, announced about a week ago.
I won’t bore you with the details of why I’m significantly uncomfortable continuing to use Cursor as my principle instrument for coding—for my main work and my various “side quests”—this blog, Musology, Kinfolx, my partner’s business apps, and the like.
Changing an IDE is like a change of “instrument”—it’s a big decision. You get used to the quirks and patterns and functions and build a (mental) muscle-memory as to where things are, and how it “ticks.” I don’t make such a decision lightly. I will no doubt, in the short term, see a hit to my productivity. As I get to know the new product, as I find the gaps (and fill them or adapt around them). But it is important to me that I stay true to my values, in that regard.
I have been eyeing off Zed Editor for a while, including the recent milestone of reaching version 1. I’d noticed that some of the dev folx I follow and admire on YouTube, like Kevin Powell, had switched. And I like the ethos, how Zed had come about, their commitment to performance and craft, and the pedigree of its founders. So Zed seemed like the logical “first port of call” for me to try, in a post-Cursor world.
This post is the first created within the Zed environment.
Cursor wasn’t just the IDE, of course. It was also the seemless integration with a bunch of frontier models, and their own “utilitarian” Composer model. So with the switch to Zed, so too a switch in how I access these models.
As I use Claude for more general LLM assistance, and I have become quite enamoured with Opus 4.x series models (and the fleeting exposure I had to Fable 5 was also positive), it would make sense to keep using that, even with a direct use of Claude Code for LLM-oriented tasks, if not within Zed’s built-in Agent support.
However, Claude’s pricing model is whack. They have USD$20 for the “standard” package. Then the next available tier is USD$100. I was paying Cursor for either their USD$20 or $60 plans (depending on the month and how active I was building). So even if we factor in the higher end of my Cursor spend, with exchange rate it’s still a hefty jump per month.
Admittedly, that would give me access to CoWork and Claude Design—both tools I’m interested in exploring. But not a 5x hit to my purse.
So, I’m trying out OpenRouter which has very transparent pricing, and local LLMs via LM Studio, as a start.
I’ve seen some commentary recently about the GLM5.2 model, and a lot of positive sentiment has been pointed towards the Qwen models. So I’m taking the opportunity to try out a few different options outside of the Anthropic sphere, while I’m at it.
It’s super early days (less than 24 hours in), and It’s way to early to judge if this new setup will work for me. I’m genuinely curious to see how it goes.
So far it seems the local LLMs, running on an admittedly lower-spec M4 MacBook Air M4, are not going to do the do (I’ve tried Google’s gemma-4-e4b, and Alibaba’s qwen3-4b-2507). GLM5.2 in the cloud seems to be quite a degree less responsive than my Cursor setup felt. I’m aware that there are more “responsive” options at higher prices available in OpenRouter. And it may be that other models perform better than GLM5.2. So even that’s a “watch this space” proposition.
At least I’m now up and running. Which is a relief TBH, as I can now withdraw my financial contribution to Musk’s empire. If this works well, I’ll make the same jump in the “day job.” But I’ll sound it out here first.
I’ll take that as a win…