Z.ai GLM

tom@tomrochette.com (Tom Rochette) — Wed, 24 Dec 2025 00:00:00 +0000

GLM 4.7 (2025-12)
- Feels a bit less smart than Claude Sonnet 4.5
  - A bit too strict on instructions following and does not appear to think about the consequence of its changes (e.g., keep __init__.py files empty, ends up deleting the code in them but not updating code that may rely on what was in those files)
- More reactive to errors rather than being proactive (e.g., will try running uv run ruff to discover that ruff isn’t installed)
- Speed is comparable to Claude Sonnet 4.5 and acceptable
- Code quality looks comparable to Claude Sonnet 4.5
- Can stay busy non-stop with the Pro plan
GLM 5.1 (2026-05)
- This is the release where I really started enjoying the model
- I stopped using Claude Sonnet/Opus (4.7-4.8 at the time): similar results, more usage, cheaper
GLM 5.2 (2026-06)
- At home I use GLM 5.2 from z.ai
- At work I use GLM 5.2 fast (provided by Fireworks) and it has been a good experience
  - Fireworks is a bit unreliable from time to time
Writing (2026-08)
- Most of my articles are now generated by GLM models, as the llm=glm-* tags on this blog show
- I pick topics I am thinking about and share those thoughts through the intermediate of an LLM generating the text
- See What the Author Brings When the Model Writes for what this means for authorship
- See My AI Workflow for how the models fit in my daily setup

z.ai on tomrochette.com