A developer’s comparison of large language models is challenging the assumption that size and broad benchmark performance reliably translate into better handling of unanswerable questions. Writing on ArrowTSX, the author points to Artificial Analysis results in which GLM-5.2 recorded a 28% hallucination rate on the AA-Omniscience benchmark, compared with 86% for GPT-5.5. The figures amount to roughly a threefold difference, but they describe one benchmark’s behavior rather than an overall error rate for either system.
The analysis defines the concern as calibration: whether a model recognizes when it lacks a valid answer instead of producing a confident response. It notes that the largest systems score strongly on Artificial Analysis’s general Intelligence Index, while GLM-5.2 reportedly comes within four points of GPT-5.5 and nine points of Anthropic’s Fable 5. The author identifies GLM-5.2 as an MIT-licensed, open-weight model with 753 billion total parameters and about 40 billion active at a time. Parameter estimates for proprietary rivals are necessarily less certain.
Other reported AA-Omniscience hallucination rates reinforce that the ranking differs from a general capability table: Anthropic’s Opus 4.8 scored 36%, Fable 5 scored 48%, and DeepSeek V4 Pro scored 94%, according to the post. On this test, lower is better. Those results should not be generalized without the benchmark’s question set, scoring rules and run conditions, and they do not establish that one model is more accurate in every task category.
The author supplemented the published scores with a single Python prompt designed around an architectural impossibility. Both GLM-5.2 and DeepSeek V4 Pro were run through OpenRouter with high reasoning effort, temperature set to 1 and the same short system instruction. GLM-5.2 reportedly identified the contradiction in about 12 seconds using roughly 800 reasoning tokens. DeepSeek spent three minutes and 26 seconds and nearly ten times as many reasoning tokens before offering a polished but incorrect solution.
That example illustrates the argument but is not a controlled, repeated evaluation. Providers, model serving and quantization can affect outputs: the post says Z.ai served GLM-5.2 and Baidu Qianfan served DeepSeek V4 Pro, both at FP8 precision. A single response can also vary across runs.
The useful conclusion is narrower than a declaration that scaling has failed. Capability, uncertainty calibration and computational cost are separate dimensions, and a model can lead on one while performing poorly on another. Buyers and developers should inspect tests aligned with their own risks, especially when a convincing wrong answer is more damaging than an explicit refusal. The reported gap makes calibration worth measuring; it does not by itself settle which architecture or model is best.
