Google reports Gemini 4 Argon abstains more and gives 2.7x to 3.9x fewer wrong answers on SimpleQA Verified and AA-Omniscience ↗
Google's safety report says Argon abstains when unsure rather than guessing, and that wrong answers on two closed-book benchmarks fell 2.7 to 3.9 times against Gemini 3.1 Pro. The figures are self-reported, and the exact rates appear only in charts. This problem's bar is a hallucination rate of 5% or below at 65% correct on AA-Omniscience.
Google DeepMind · Gemini 4 Argon Safety & Security Report