GPT-5.5 hallucinates 3x more than MIT-licensed GLM-5.2 (arrowtsx.dev)

Recent evaluations reveal that OpenAI's GPT-5.5 exhibits a staggering 86% hallucination rate, significantly higher than Z.ai's GLM-5.2, which suffered a 28% hallucination score despite being only 753 billion parameters compared to an estimated 1-2 trillion for GPT-5.5. This discrepancy highlights a crucial shift in the AI landscape, where larger models do not necessarily translate to better performance or reliability. The findings underscore an emerging skepticism within the AI community regarding the efficacy of simply scaling model parameters and training datasets. As AI research pivots towards quality over quantity, the performance of these models raises critical questions about training methodologies and the importance of uncertainty calibration. Z.ai's GLM-5.2 excelled in complex reasoning tasks, identifying logical fallacies with remarkable efficiency, while larger counterparts like DeepSeek V4 Pro faltered by confidently generating incorrect responses after extensive computational effort. This revelation suggests that the future of AI development must balance raw capability, reduction of hallucination rates, and computational efficiency, as the overemphasis on model size risks diminishing true intelligence.