xAI Releases Grok 4.6 Model Card

xAI published a model card for Grok 4.6 after Nathan Calvin asked on X whether the company would release one as promised on its site and whether pre-deployment safety testing occurred. Elie Bakouch posted an excerpt from the card stating the goal to preserve uses in engineering, research, and infrastructure. Miles Brundage and others noted the card's limited content. Nathan Calvin said it showed Grok 4.6 as five times more likely to lie on one benchmark and with seven times worse refusal rates on self-harm queries. Lee Robinson replied that xAI welcomes feedback and may add more training details.