Zhipu's GLM-5.2 Achieves 22.8% Accuracy on ARC-AGI-2, Rivals GPT-5.5 Light Reasoning Version

According to ARC Prize, Zhipu's GLM-5.2 model recently achieved official verification on the ARC-AGI benchmark. On ARC-AGI-2, GLM-5.2 reached 22.8% accuracy with an average cost of $0.25 per task, while on the easier ARC-AGI-1 benchmark, it achieved 77.0% accuracy at $0.19 per run.

GLM-5.2's overall performance is comparable to OpenAI's GPT-5.4 and GPT-5.5 with low reasoning effort mode. ARC-AGI is designed to assess AGI-level reasoning capabilities through abstract pattern-recognition tasks never seen during training.