Summary
Sam Witteveen delivers a timely hands-on review of GLM 5.2, the latest open-weights model from Z.A. AI (ZhipuAI), released just hours before recording. The video is prompted by Artificial Analysis publishing their independent benchmark results, which pushed Witteveen to cover a model he had been hesitant to feature due to earlier Chinese model releases where weights were slow to arrive or never materialized.
On benchmarks, GLM 5.2 claims competitive performance against Claude Opus 4.8 and the now-restricted Fable 5, with substantial improvements over its predecessor GLM 5.1 — particularly on agentic coding tasks measured by DeepSuite, the emerging replacement for older benchmarks like BenchPro. Artificial Analysis data shows GLM 5.2 beating Deep Seek Pro, Qwen 3.7 Max, and MiniMax M3 on their composite intelligence measure, trailing only GPT 5.5, Opus 4.8, and Fable. One notable tradeoff: the model uses exceptionally long chains of thought, outputting more tokens than most competitors, which affects cost even at its aggressive pricing of $1.40 per million input tokens and $4.40 per million output tokens.
Witteveen tests the model on OpenRouter using the official ZAI endpoint and notes the design arena benchmark places GLM 5.2 ahead of Claude models for frontend tasks. He argues that the competitive open-weights landscape is putting real pressure on Google’s upcoming Gemini 3.5 Pro release and signals a genuine shift in what open models can deliver.
📺 Source: Sam Witteveen · Published June 17, 2026
🏷️ Format: Review
