# Trust does not carry forward > Source: > Published: 2026-11-05 22:25:17+00:00 DeepSeek V4 Pro runs 1.6 trillion parameters, 49 billion of them active, and scores 44 on the Artificial Analysis Intelligence Index. On the AA-Omniscience benchmark it posted a 94% hallucination rate. When it hit a question it could not answer, it said so about 6% of the time and confidently invented the rest. Line that up against the rest of the field and it reads as a pattern rather than an outlier. GLM-5.2 scored 28%, Opus 4.8 36%, Fable 5 48%, GPT-5.5 86%. Hallucination rate does not track parameter count. The two worst at admitting they do not know are GPT-5.5 and DeepSeek V4 Pro, both huge models the author picks out as the clearest hallucination leaders. Major labs are getting skeptical of pure scaling, and the ceiling on the old assumption became visible when Claude Fable 5 was restricted by the US government three days after release, the first US AI ban on national-security grounds, because a single jailbreak was judged too much risk. One of the largest models in the world was pulled over one failure mode. The capability case for scale is also thinner than it looks. GLM-5.2 is Z.ai's newest release: 753 billion parameters, roughly 40 billion active, MIT-licensed open weights. On the Intelligence Index it lands within 4 points of GPT-5.5 and 9 of Fable 5. GPT-5.5 is proprietary and, by the author's estimate, sits in the 1 to 2 trillion parameter range, conservatively 1.5 to 2 times larger than GLM-5.2. An open-weight model getting that close says raw intelligence has plateaued. The biggest models still win the benchmark on paper. On whether they tell you when they are out of their depth, the extra size buys nothing and sometimes costs you. The author tested this with one Python question built around a clear architectural flaw: a single-threaded task asked to perform multiplexed I/O without ever yielding or using system polling. It cannot be done as stated. GLM-5.2 recognized the impossibility in about 12 seconds and 800 reasoning tokens. DeepSeek V4 Pro spent 3 minutes and 26 seconds in a reasoning loop, burned close to ten times the reasoning tokens, and returned a cleanly structured, confidently wrong solution that treated the problem as solvable. Both ran at "high" reasoning effort, temperature 1, on OpenRouter, under the same system prompt, at FP8 precision, GLM-5.2 served by Z.ai and DeepSeek V4 Pro by Baidu Qianfan. The author's explanation is that training on large volumes of factual, non-theoretical data teaches a model to always have an answer ready. Scale the corpus and the parameter count and confidence grows faster than correctness. The model learns the shape of a correct-sounding response and reaches for it even when the honest output is a refusal. For anyone putting a model into a workflow, that changes the selection problem. A high Intelligence Index score sitting next to an 86% hallucination rate describes a model that becomes a liability anywhere a wrong answer costs more than no answer: automated code review, anything that feeds a downstream system without a human checking each step, any pipeline that trusts output because it looks finished. The confidently wrong DeepSeek solution is the failure that ships, precisely because it reads as complete. So read calibration numbers next to capability numbers. AA-Omniscience, or any measure of how often a model refuses versus fabricates, belongs beside the Intelligence Index in an evaluation, and a model that refuses when uncertain is worth more in production than one scoring two points higher that makes things up. The author frames it as a trilemma the industry has not solved: raw capability, uncertainty calibration, and computational efficiency. DeepSeek V4 Pro lost two of the three in a single question.