Zhipu AI's GLM-5.2 becomes best open model, trails frontier by months
New open model beats all peers but lacks vision and costs too much for bulk tasks.
GLM-5.2, released last week by Chinese lab Zhipu AI, claims the title of best open model with impressive benchmarks and a significant step up from GLM-5.1. According to LessWrong blogger Zvi, it's closer to the frontier than DeepSeek R1 was during its peak, yet still 4–7 months behind closed models like Claude or GPT. The model is distilled from Claude and excels in core text tasks, matching Opus 4.7 in many areas.
However, GLM-5.2 suffers from critical limitations: lack of vision support, practical difficulties in deployment, and a pricing sweet spot that's neither cheap enough for bulk processing nor capable enough for top-tier tasks. Zvi notes its best use case is for those who inherently need openness, but broader enterprise adoption remains elusive. The model represents progress but doesn't yet shift the competitive landscape.
- GLM-5.2 is the strongest open model, outperforming all previous open alternatives including DeepSeek R1.
- It is distilled from Claude and trails closed frontier models by 4–7 months in capability.
- Lacks vision and other features, making it a niche choice for tasks requiring openness over raw power.
Why It Matters
Open-source AI narrows the gap but still can't match closed models for enterprise-grade vision and cost efficiency.