Key Points
- 1.Nate Herk tests GLM 5.2 inside Claude Code and finds it fast, much cheaper, and a clean fit for the Claude Code harness.
- 2.On a website-design task GLM 5.2 produced comparable output to Opus 4.8 in about 4 minutes versus roughly 15, at about 5x lower cost per token.
- 3.Opus 4.8 stayed more precise on a subtle edge case (duplicate records with values like true vs 1), so it is still the stronger model overall.
- 4.The practical takeaway: route roughly 80% of knowledge work to a cheaper model like GLM 5.2 and reserve Opus for the 10-20% that needs heavy reasoning.
- 5.He runs it with a 1M-token context on a 60-dollar-per-month Z.AI plan and shows the full setup inside Claude Code.
Summary
GLM 5.2 vs Opus on speed and cost
Across design and build prompts, GLM 5.2 finished some tasks far faster than Opus (about 4 minutes vs 15 on a website design) at roughly 5x cheaper per token, with similar output quality. On the heaviest-reasoning tasks, Opus was faster.
Where Opus still wins
A judged head-to-head found Opus 4.8 more precise on a subtle edge case involving duplicate records with values like true vs 1. GLM 5.2 was solid but less exact on the hardest details.
Match the model to the task
The core skill going forward is picking the right model per task: about 80% of knowledge work runs fine on a cheaper model like GLM 5.2, with a top-tier model like Opus reserved for the 10-20% that needs deep reasoning.
Setup and cost
He runs GLM 5.2 with a 1M-token context on a 60-dollar-per-month Z.AI plan and walks through getting it connected inside Claude Code, including billing and quota usage.
Worth watching for
Developers and AI builders deciding when to use a cheaper model in Claude Code versus a top-tier model like Opus.
- claude-code
- glm
- models
- cost