Summary
GLM 5.2, the open-source 753-billion-parameter model from Zhipu AI, is drawing serious attention from AI developers after creator Nate Herk spent a full day stress-testing it inside Claude Code. In this hands-on walkthrough, Herk shows exactly how to configure GLM 5.2 as the underlying model in Claude Code and runs it against Claude Opus 4.8 across real tasks — automated video editing, single-prompt website design, and coding assignments judged by a third-party model.
The benchmark numbers are compelling: a website design task wrapped up in 3 minutes 59 seconds with GLM 5.2 versus 14 minutes 59 seconds with Opus, at roughly one-fifth the cost. GLM 5.2 pricing via API comes in at $1.40 input / $4.40 output per million tokens, compared to Opus 4.8’s $5 / $25. Herk accesses it through Z.AI on a $60/month subscription and burned through roughly 357,000 tokens editing a single 23-second video clip. The model’s 1-million-token context window is a key enabler for long agentic sessions.
The core argument is strategic model routing: GLM 5.2 handles the bulk of coding, design, and research tasks with speed and cost efficiency, but Opus still wins on tasks requiring precise reasoning — like catching edge cases involving duplicate records with type mismatches. Herk’s takeaway for AI power users is that knowing which model to deploy at each step of a workflow is becoming as important as knowing how to prompt — and defaulting to the most powerful model for everything is increasingly hard to justify on cost alone.
📺 Source: Nate Herk | AI Automation · Published June 19, 2026
🏷️ Format: Tutorial Demo
