Grok 4.1: xAI's November 2025 Model for grok.com and X, Thinking Modes, Training and Evaluations

What Grok 4.1 is

Grok 4.1 is a model from xAI, now part of SpaceXSI, released on 17 November 2025 to all users on grok.com, X and the iOS and Android apps. It rolled out immediately in Auto mode and could be chosen explicitly as Grok 4.1 in the model picker. xAI said it brought significant improvements to Grok's real-world usability, calling it exceptionally capable in creative, emotional and collaborative interactions, more perceptive to nuanced intent and more coherent in personality, while keeping the intelligence and reliability of its predecessors. It is a proprietary model with no published weights. On LMArena's text leaderboard it has 1459 Elo (rank 64).

Thinking and non-thinking

Grok 4.1 comes in two configurations. Grok 4.1 Thinking, code-named quasarflux, uses reasoning tokens before it answers; the non-reasoning configuration, code-named tensor, uses no thinking tokens and responds immediately. This page covers the Thinking configuration. At launch xAI offered Grok 4.1 in its consumer apps; for developers, xAI released the separate Grok 4.1 Fast models on its API two days later.

How xAI trained it

xAI says it used the same large-scale reinforcement learning infrastructure that powered Grok 4 and applied it to the style, personality, helpfulness and alignment of the model. To optimise these non-verifiable reward signals, xAI developed methods that use frontier agentic reasoning models as reward models, evaluating and iterating on responses autonomously at scale. Between 1 and 14 November 2025, xAI ran a silent rollout of preliminary Grok 4.1 builds to a growing share of traffic on grok.com, X and the mobile apps, with continuous blind pairwise evaluations; it reports that users preferred Grok 4.1 over the previous production model 64.78% of the time.

Evaluations

xAI evaluated Grok 4.1's emotional intelligence on EQ-Bench3, an LLM-judged test of 45 roleplay scenarios, run with the default sampling parameters, the prescribed judge and no system prompt. It also measured creative writing on Creative Writing v3, which uses 32 writing prompts across 3 iterations. For hallucinations, xAI focused post-training on information-seeking prompts and reports significant reductions in the hallucination rate on a stratified sample of real production queries, and on FActScore, a public benchmark of 500 biography questions, both measured with the non-reasoning model using web search tools. These are xAI's own evaluations, and xAI publishes a Grok 4.1 model card. xAI reports both rubric scores and normalised pairwise scores for EQ-Bench3 and Creative Writing v3, computed with the official benchmark repositories, and said it was working with the benchmark authors to get its results onto their leaderboards.

Availability

xAI's models documentation lists Grok 4.1 as released to all users on grok.com, X and the iOS and Android apps, with no separate API price for this configuration. xAI has since released later Grok models, including Grok 4.3. Its launch post gives side-by-side examples comparing Grok 4.1's replies with the previous Grok on emotional and creative writing prompts, along with the methodology for each evaluation.

Official Resources