GLM-5.2 posts 51 on Artificial Analysis Intelligence Index v4.1, first open-weights model above 50

Z.ai released GLM-5.2 with unchanged 744B/40B MoE parameters yet extended context to 1M tokens. The model records gains concentrated in scientific reasoning, lifting CritPt to 21 percent and HLE to 40 percent, alongside TerminalBench v2.1 at 78 percent. Output token consumption rises to 43k per task, of which 37k are internal reasoning traces.

On GDPval-AA v2 the score reaches 1524, placing GLM-5.2 level with GPT-5.5 xhigh and ahead of all prior open-weights entries. Cost per task settles at 0.46 dollars, locating the model on the Intelligence-Cost Pareto frontier despite higher absolute token counts than DeepSeek V4 Pro. MIT license and multi-provider inference availability accelerate downstream fine-tuning and distillation pipelines.

The result marks the first independent benchmark leadership by a fully open-weights system above the 50-point threshold. Closed labs retain edges in raw latency and calibration, yet reproducible weights plus documented training details compress iteration cycles for academic and startup labs. Subsequent releases from the same series are expected to target token efficiency while preserving the observed reasoning depth.