GLM 5.2, the open-weight coding model from Z.AI, completed four production tasks for $3.36 across roughly 6 million tokens. Those tasks: a codebase architecture audit on a live Next.js app, an HTML roadmap page, a landing page redesign that matched an existing design system on the first attempt, and a 45-minute autonomous bug-hunting session ingesting both Sentry errors and Vercel logs. The output was a prioritized bug-fix dashboard the author is shipping from.
The case for reading this in full is not the price. It is the setup details. The author walks through exactly how to connect GLM 5.2 to both Cursor and Claude Code via OpenRouter, where the model stumbled, and what a 45-minute unattended agentic run actually looks like against real production logs. That failure section matters as much as the wins.
The comparison benchmark here is Claude Opus. If a $3.36 open-weight model can displace Opus on autonomous coding tasks, the cost structure of AI-assisted development shifts significantly. The original covers the full token cost breakdown, the specific prompts used, and where GLM 5.2 lost coherence under pressure.
[READ ORIGINAL →]