Grok 4.1 Thinking

1, bringing significant improvements to the real-world usability of Grok. The model is exceptionally capable in creative, emotional, and collaborative interactions.

It is more perceptive to nuanced intent, compelling to speak with, and coherent in personality, while fully retaining the razor-sharp intelligence and reliability of its predecessors. 1 Thinking uses thinking tokens for deeper reasoning, achieving state-of-the-art performance on blind human preference evaluations.

The model features reduced hallucinations compared to previous versions, with significant improvements in factual accuracy for information-seeking prompts.

To achieve this, xAI used large scale reinforcement learning infrastructure to optimize style, personality, helpfulness, and alignment, developing new methods that use frontier agentic reasoning models as reward models to autonomously evaluate and iterate on responses at scale.

1 Thinking includes comprehensive safety mitigations including refusal training, input filters for restricted knowledge, and adversarial robustness measures.

The model demonstrates strong refusal rates on harmful queries (7% answer rate on chat refusals, 14% on agentic refusals) and improved honesty training to reduce deception.