I’ve been trying to get off the frontier labs for a while now, for reasons I got into a bit in “good enough” models. The short version is that the cheap models have gotten good enough that paying frontier prices for mechanical coding tasks feels silly. My first thought was to try out the open weights models that have been coming out recently, but this turned out to be harder than I’d expected. Not because of the models themselves, but because finding somewhere decent to access them is surprisingly complicated. What follows is a haphazard chronicle of roughly a month of my poking around at various coding subscriptions, and how each poke has worked out so far.

The landscape of available models is changing super fast right now: as I write this Anthropic’s Fable is the undisputed best for sheer intelligence, but is also painfully expensive. OpenAI’s Sol is very close behind and much more reasonably priced, but still prohibitively expensive if you’re not on one of their subscription plans. At this point raw intelligence is no longer the only thing I care about though. Even relatively weak models (compared to the frontier) such as Zai’s GLM 5.2 are quite intelligent and more than enough for many coding tasks, at a way lower price.

An idea I’ve been trying to articulate recently is that the hard part about working with AI in 2026 isn’t actually getting them to do what we asked. Anyone who’s been using coding agents in 2026 knows this. In early 2025 your time was spent fixing up misunderstandings and mistakes the AI made. In mid 2026, that’s no longer the case. The agents are pretty good at doing what you ask them to do. The hard part today is figuring out how to accurately describe what you intended. The class of errors most worth worrying about today looks more like: “the agent understood me perfectly, followed my instructions to the letter, and in doing so showed me exactly how wrong my instructions were.”

PACU is an acronym I’ve started using with LLMs, short for “Pause and Check Understanding.” The key is it forces your agent to regurgitate its interpretation of your prompt back to you. If you spend a lot of time using AI coding tools you probably have your own variant in your AGENTS.md:

A weekend vibe project I’ve been spending a lot of time on lately has turned out to be more useful than expected in practice, so I wanted to share a bit about what it is, why it works, and why I like it.

Here’s a snapshot of my AI coding workflow as of April 2026. While there’s a good chance this will be obsolete by June because two months is an eternity in AI time, now also seems like a good time to write it down because for once it hasn’t changed significantly in months.