Grok 4.6 posts 34.3% AA-Omniscience hallucination versus 92.2% for GPT-5.6 Sol

SpaceXAI’s Grok 4.6 records a 34.3% hallucination rate on Artificial Analysis’s AA-Omniscience test, far below GPT-5.6 Sol’s 92.2%, even as the two models tie on the lab’s broader Intelligence Index — a split that is sharpening debate over whether raw capability scores capture reliability in production. Independent trackers describe the August 12, 2026 release as the first 2026 flagship whose launch cut hallucination rather than buying accuracy with more fabrication.