TL;DR
The short version
On How I AI, Claire Vo wired GLM 5.2 — an open-weight model from Beijing lab Z.ai — into Cursor and Claude Code and ran it on her own production codebase. It handled codebase comprehension and design well, stumbled on React, and cost $3.36 for the whole session.
The reusable point for anyone running a team: the question is no longer 'is it as smart as Opus' but 'is it smart enough for this task, at this price' — which makes model routing, not model loyalty, the operator skill.
Built on Claire Vo's GLM 5.2 review on How I AI. Cost and PR figures are self-reported from her dashboards; benchmark standings are unverified.
The default assumption was wrong
Most teams route everything to the most expensive model by reflex — Opus or GPT 5.5 for every task, because reasoning is the bottleneck and you don't want to think about it.
Vo's test breaks that reflex. She self-hosted GLM 5.2 — open-weight means the trained weights are public, so you run it on your own inference instead of paying a frontier API — and pointed it at her own ChatPRD codebase. Three real tasks, no rehearsal, for $3.36 across roughly 6 million tokens. Against Opus pricing, that is a rounding error.
The cost isn't the interesting part. The interesting part is that the work was good.
Good enough is a moving line
She gave it three jobs. It explored an unfamiliar Next.js codebase and explained the architecture fast and accurately. It redesigned a marketing hero section against an existing design system — the harder test, because it has to match house style, not invent something new — and matched it, even reproducing the brand color other models miss. And it ran a 45-minute autonomous job: pull 72 hours of Sentry and Vercel errors, build a prioritized fix plan, ship it in a clean canvas.
The honest weak spot was front-end generation. On that last task it queried data and reasoned well but stalled repeatedly on TypeScript and JSX.
I don't think it can write React, which is 98% of what I do with these models.
Claire Vo, How I AI
So the verdict isn't 'it replaces Opus.' It's narrower and more useful: strong on comprehension, design, and tool-heavy backend work; weak on React. A model with good taste on cheap tokens is enough for a real slice of daily work.
Route by task, not by loyalty
If one cheap model can do design and codebase work at near-frontier quality, paying frontier prices for those tasks is a leak. The skill isn't picking the smartest model — it's matching task to model and saving the expensive one for the jobs that need it.
Open weights make that routing durable a second way: you're not locked to one vendor's terms. If a provider changes pricing or access, you swap the inference layer underneath and keep working — optionality worth as much as the cost savings to any team that's been burned by a deprecation.
The minimum viable test
You don't need to trust a vibe check. The setup is three moves: pick a provider (Vo used OpenRouter, a gateway that fronts many models), swap in its API key, and route your model calls to the new model string. The same pattern works for Cursor, Claude Code, and Codex. One gotcha — Cursor's base-URL override has to end in `/cursor`, which is undocumented.
Then run your own three tasks on a codebase you know: comprehension, a design-system match, and one long autonomous job. An afternoon and a few dollars tells you which of your daily tasks no longer need the premium model.
Treat the specifics as signals, not settled facts — the benchmark standings weren't traced to a primary source and the cost came off Vo's own dashboard. The reusable point survives the caveats: the cheapest way to find your routing line is to test it on work you can already judge.
