Zhipu AI's latest MoE (Mixture of Experts) flagship model, GLM-5.2, just scored 51 points on the Artificial Analysis Smart Index v4.1 — beating out MiniMax-M3 (44 points), DeepSeek V4 Pro (max, 44), and Kimi K2.6 (43) to claim the top spot among open-source models worldwide.
https://twitter.com/ArtificialAnlys/status/2067135640249209175
On the GDPval-AA v2 benchmark, which simulates real-world knowledge work, GLM-5.2 racked up 1,524 points — a human baseline is 1,000 — outscoring MiniMax-M3 (1,418) and DeepSeek V4 Pro (max, 1,328), and landing right alongside the closed-source frontier model GPT-5.5 (with xhigh reasoning enabled). Compared to its predecessor GLM-5.1, scientific reasoning on CritPt jumped 16 percentage points to 21%, HLE went up 12 points to 40%, TerminalBench v2.1 climbed 16 points to 78%, and GPQA Diamond hit 89%.
GLM-5.2 also sits at a sweet spot on the intelligence-to-cost Pareto frontier. Because each task now averages 43,000 output tokens (up from 26,000 for GLM-5.1), a single task costs about $0.46 — higher than GLM-5.1's $0.25 and DeepSeek V4 Pro (max)'s $0.05, but still way cheaper than closed models of similar smarts.
The model packs 744 billion total parameters with 40 billion activated per inference, expands its context window from 200K to 1 million tokens, and is open-sourced under the MIT license. You can try it now via Zhipu's official API (priced at 1.4 yuan per million input tokens and 4.4 yuan per million output tokens), or through platforms like SiliconFlow, DeepInfra, and Nebius AI.