Independent Analysis: Grok 4.5 Improves Accuracy but Doubles Its Hallucination Rate

My take: Artificial Analysis, an independent benchmark source, published its full analysis of Grok 4.5 today: the model's accuracy rose from 35% to 52% compared to Grok 4.3, but the hallucination rate doubled, from 25% to 54%. The model is right more often, but when it is wrong, it is now more confident about it.

The distinction matters in practice. For code generation, plan structuring, or creative tasks, a more accurate model has real value. For research, factual analysis, or any work where errors have consequences, a model that hallucinates with greater conviction is a higher risk than one that is more cautious. Grok 4.5's competitive price ($2 per million input tokens) and its agentic performance make it attractive for certain workflows; but the benchmarks xAI publishes comparing itself to competitors deserve a critical read, since the company is both judge and contestant in its own tests. Independent numbers are the ones that matter.

AI does not make you smarter: it makes you more efficient. But that efficiency depends on knowing which model to assign to each task type. Does your team have a defined policy for choosing the right model based on the factual reliability each task requires?