GPT-5.5 hallucinates 3x more than MIT-licensed GLM-5.2
Points and comments are a snapshot, not live.
Larger models like GPT-5.5 hallucinate more than smaller MIT-licensed GLM-5.2.
A blog post argues that scaling model size no longer yields proportional intelligence gains, citing hallucination rates on the AA-Omniscience benchmark: GLM-5.2 at 28%, Opus 4.8 at 36%, Fable 5 at 48%, GPT-5.5 at 86%. DeepSeek V4 Pro scored 94%. The author describes a trilemma among raw capability, uncertainty calibration, and computational efficiency. A test with a Python question showed DeepSeek V4 Pro producing a confidently incorrect solution after extended reasoning, while GLM-5.2 quickly identified the impossibility.
What commenters are saying
Commenters debated the interpretation of hallucination rates. One noted the metric is conditional on the model not knowing the answer, making absolute rates different: Opus 4.8 had 19% absolute hallucination vs GLM-5.2's 21%. Others argued that model size alone doesn't explain differences, citing DeepSeek-V4 Flash as smaller but top-ranked. A minority criticized the focus on prompting to mitigate hallucinations, calling it a version of 'you're holding it wrong.' Another commenter distinguished between knowingly bullshitting and confidently misremembering.