GLM 5.2 Tops Open Models on ARC-AGI Benchmarks

GLM 5.2 has achieved the highest scores for an open-weight model on the ARC-AGI benchmarks. It scored 77.0% on ARC-AGI-1 and 22.8% on ARC-AGI-2, according to the ARC Prize's verified leaderboard. This model uses chain-of-thought reasoning. What's interesting is that ARC-AGI tests fluid intelligence, focusing on inferring novel rules rather than memorization. While GLM 5.2 leads open models, its 22.8% on ARC-AGI-2 is still far behind frontier models like Gemini 3.1 Pro, which scores 77.1%. This gap highlights a significant difference in handling compositional generalization. The bottom line is that while GLM 5.2 represents a milestone for open models, general reasoning at scale remains a challenge compared to well-resourced frontier labs. This shows us the current state of open models in relation to advanced AI tasks.