\n\n
Recent benchmark updates reveal that hallucination rates across leading AI models are climbing. OpenAI’s latest GPT‑6 Astra, released in early September, cut its hallucination rate to 51.3% at peak but still lagged far behind older versions, while the median hallucination rate for public models jumped from 48.4% to 55.2% between the first half and the July‑September 2026 period. Health‑focused tests show GPT‑5’s “thinking” mode hallucinated on just 1.6% of hard cases, yet newer OpenAI variants such as o3 and GPT‑4o still produce hallucinations in 33%‑16% of queries. Legal domains are especially affected, with AI‑generated false material surfacing in 2,149 court filings by early October. Meanwhile, the Vectara leaderboard highlights a modest low point—antgroup’s finix_s1_32b at 1.8%—but most flagship models sit between 7% and 13% for summary hallucinations. These trends underscore growing reliability gaps as AI systems become more powerful, prompting urgent calls for better verification, multi‑model checks, and domain‑specific mitigations. [[1, 2]]