The new HalluHard benchmark, developed by researchers in Switzerland and Germany, shows that large language models remain unreliable. According to the study, even the strongest model tested, Claude Opus 4.5, still produced hallucinated content in around one third of cases when web search was enabled. Without web search, that rate doubled. GPT-5.2 Thinking also recorded a hallucination rate of 38.2% with web search enabled. This suggests that web search can reduce hallucinations, but cannot eliminate them entirely.
HalluHard’s strength lies in its realistic evaluation approach. The benchmark is based on simulated conversations designed to reflect multi-step interactions. An integrated algorithm systematically analyses all sources and files referenced by the model, including full texts and PDFs. The aim is to identify statements or details that are not supported by the cited material. This makes it possible to detect hallucinations that may go unnoticed in more superficial evaluations.
By using multi-step dialogues, HalluHard differs from many previous benchmarks. These often rely on single-turn interactions, offer only limited evaluation options and can also reach a performance ceiling relatively quickly.
By Ida Kandler
Source: Linux Magazin, 11 February 2026, by Jens-Christoph Brendel
![Large AI models still hallucinate in more than 30% of cases 1 © [2025] [𝘒𝘐-𝘎𝘦𝘯.] 𝘔𝘢𝘳𝘤𝘦𝘭 𝘖𝘩𝘳𝘦𝘯𝘴𝘤𝘩𝘢𝘭𝘭. 𝘉𝘢𝘴𝘪𝘦𝘳𝘦𝘯𝘥 𝘢𝘶𝘧 𝘞𝘦𝘳𝘬𝘦𝘯 𝘷𝘰𝘯 𝘈𝘯𝘥𝘳𝘦𝘢𝘴 𝘖𝘩𝘳𝘦𝘯𝘴𝘤𝘩𝘢𝘭𝘭.](https://research-ki.de/wp-content/uploads/2026/02/flying_elephant_over_a_city_skyline_ink_drawing_detail_5e2103a6-3b4d-4643-b114-bbf1fffdd111_2.png)