The quest for AI tooling nerdvana took a little turn this past month.
For a few months, I was happy using Cursor – one of several VS Codium clones in the IDE space with a customized plugin. It had high DWIM-potential factor – I was confident what would happen with a given prompt, it stayed within its bounds and didn’t cost much at all (~$24/month, IIRC, with the most I ever consumed being 75% quota).
But then the company got bought-out and I didn’t want to be funding the new owners, for obvious reasons. Plus it started trying to foist various Grok models on me, whose output I’m not prepared to countenance using.
After some exploring, I discovered Zed – open-source, a very nice IDE, offering a choice of Codium / Emacs / Vim keyboard shortcuts at startup, a rich plugin and theme ecosystem and a similar layout to Cursor (except for some reason it puts the agent chat window on the left and project directory-tree on the right…). So Zed has become my new home IDE of choice for the foreseeable. Days were when I actually used emacs, not merely emulated it… and now I don’t do that either.
One of the things I learned through the Zed approach was MCP, a system of storing remote API endpoints and their credentials for later invocation (don’t-call-me-“SOAP”-,honest). That alone has proved radically enjoyable, using it to inspect, query and update a WordPress server from the comfort of my IDE. I could get used to this way of working…
After Cursor, for the past month, I experimented with using Qwen3.8 running locally on the new macbook pro, with a native Claude subscription as a fallback. Where Cursor had been a middle-of-the-road approach, paying a little for everything, this split my model selection into two, leaving me to tread a line between free and expensive.
Qwen3.8 is a 30b-class model. In my experience, it’s very accurate – high DWIM-potential, easy to craft a prompt for – can generally be relied upon to stay within its bounds and, while definitely on the slow side, it punches above its weight with some really quite sizeable jobs returning solid data eventually. (“Search for the top 7 CSS & JS frameworks, download their documentation and build a crosswalk table of corresponding classes, widgets or block structures where possible” – what a prompt, what a resultant analysis!) So: free, check; accurate: check; impressive: check.
By contrast, Claude has barely lasted the month. I had somewhat of an averse reaction to it:
At $18/month for basic subscription and extra for the tokens it got ludicrously expensive – with the one month’s subscription and a few top-ups, I hit around £75 for a fortnight’s usage.
Claude Sonnet (even v5) had negative DWIM-potential. I spent an hour going round and round in circles getting it to admit its own bugs and omissions – that wilful misunderstanding seemed to be the norm across several prompts and projects.
Claude Opus (v5) was marginally better at doing what I said, and I’ve seen it do some seriously impressive work (16 commits on a shared project at work, each with excessively long commit message, implementing radically different concepts), but it has no boundaries at all. I’d tell it to edit a script to do something and it would go off on its own, not knowing the difference between coding, testing and outright trying to run the thing, complaining about its own sandbox, complaining about my private LANs, trying to run sudo to edit /etc/hosts… That’s not powerful. That’s just waaayyyyy so far out of control it couldn’t touch bottom with a long stick.
So, Claude hit the dustbin with considerable prejudice. Expensive, inaccurate, frustrating, incompetent. Biggest FAIL yet.
So I spent the past weekend looking at alternatives. Apparently prompt “routing” is a thing – there are forum discussions discussing how best to choose models by type of prompt being run. I don’t have the time for that – least of all for constructing regexps for my various thinking, tool-calling and coding purposes – but I did want a feel similar to Cursor under my fingers again.
So – using “AI” – I searched on duckduckgo for “LLM inference aggregator” and came up with a couple – OpenRouter and Ofox. Of these, ofox seems to have the slightly clearer pricing and are up-front about the 128 models they run and how much they cost per M-token in & out. Wow, lots of data.
With a couple of prompts – using “AI” – Qwen3.8 downloaded Ofox’s pricing page, converted it to TSV, added columns for zscores[0] of the costs, context window size and updated-date fields, summed the zscores to a final score and gave me a file I could open in LibreOffice. There was indeed a value gradient – the summed zscores (bang-for-buck) was quite a range from 1 to -2.5 or so. Less surprisingly, it was the Clade v5 models – Sonnet, Opus and particularly Fable – that came absolutely last down the pile. So I selected about 10 models, spanning the range from good value (but better compute than Qwen3.8 running locally – the best seemed to be z.ai’s GLM-5.2-Flash model) down to the Claude options.
Finally, I used Qwen3.8 running locally yet again to edit Zed’s settings.json config file, adding each of the models I’d selected to the ofox provider section in bulk.
The amusing thing is how much “AI” has been useful. I could’ve done everything myself, probably using awk rather than python even for maximum hackish-kludge appeal, but it’s so much easier just to spec-out what you want, wave a paw at an LLM and come back 10mins later. It’s a geekish trait, of course – appreciation for recursion dates back to days when a language’s maturity was measured by whether its compiler could be written in its own language or not. I wonder if the using-AI-to-configure-AI thing heralds the start of a new era…
And so far I’ve done a tonne of work using GLM-5.2-Flash – editing the WordPress website, analyzing colo-server web-server logs and syslogs for reasons why the database server was crashing at irregular intervals, quality-assessing 3.5MiB of Docbook documentation – all for under $1 for 4 days.
For now, nerdvana is achieved.
[0] “number of standard-deviations away from the mean”, a way to reduce any parameter to unit-less number regardless of scales and offsets. Far better to have something cardinal, suited to adding together into cumulative scores, than try to kludge together a mental picture of “desirability” by ordinal positions sorting by one column after another.