TL;DR

xAI just launched Grok 4.1 – a major upgrade that now ranks #1 on LMSYS Text Arena (1483 Elo with reasoning), dominates emotional intelligence and creative writing benchmarks, reduces hallucinations dramatically, and was preferred by real users 64.78% of the time over the previous Grok version. It’s rolling out today to all users on grok.com, X, iOS, and Android.

Key Takeaways

• Grok 4.1 (Thinking mode, codename “quasarflux”) achieves #1 on LMSYS Text Arena with 1483 Elo – 31 points ahead of the best non-xAI model.
• Even the non-reasoning “fast” version (codename “tensor”) ranks #2 globally at 1465 Elo, beating every other model’s full-reasoning score.
• Tops EQ-Bench3 emotional intelligence leaderboard and Creative Writing v3 benchmark.
• User preference win rate of 64.78% vs previous Grok during two-week silent rollout.
• Hallucination rate dropped from ~12% → 4.22% on real-world info-seeking queries.
• Trained using massive RL infrastructure plus new frontier agentic models as autonomous reward judges.
• Available right now in Auto mode and selectable as “Grok 4.1” in the model picker.

Detailed Summary of the Grok 4.1 Announcement

On November 17, 2025, xAI released Grok 4.1, calling it a significant leap in real-world usability. While raw intelligence remains on par with Grok 4, the focus of 4.1 is personality, emotional depth, creativity, coherence, and factual reliability.

The model was refined using the same large-scale reinforcement learning pipeline that powered Grok 4, but with new techniques that allow frontier-level agentic reasoning models to autonomously evaluate subjective rewards (style, empathy, nuance) at massive scale.

A two-week silent rollout (Nov 1–14) gradually exposed preliminary builds to increasing production traffic. Blind pairwise evaluations on live users showed Grok 4.1 winning 64.78% of comparisons.

Benchmark Dominance

• LMSYS Text Arena: #1 overall (1483 Elo Thinking), #2 non-thinking (1465 Elo)
• EQ-Bench3: Highest emotional intelligence Elo (normalized)
• Creative Writing v3: Highest normalized Elo
• Hallucinations: Reduced from 12.09% → 4.22% on production queries; FActScore error rate from 9.89% → 2.97%

The announcement includes side-by-side examples (grief over a lost pet, creative X posts from a newly-conscious AI, travel recommendations) where Grok 4.1 sounds dramatically more human, empathetic, and engaging than previous versions or competitors.

My Thoughts on Grok 4.1

This release is fascinating because xAI is openly prioritizing the “feel” of the model over pure benchmark-chasing on math or coding. Most labs still focus on reasoning chains and MMLU-style scores, but xAI just proved you can push emotional intelligence, personality coherence, and factual grounding at the same time — and users love it (64.78% preference is huge in blind tests).

The fact that the non-reasoning version already beats every other company’s best reasoning model on LMSYS suggests the base capability is extremely strong, and the RL alignment work is doing something special.

Reducing hallucinations by ~65% on real traffic while keeping responses fast and natural is probably the most underrated part of this release. Fast models with search tools have historically been the leakiest when it comes to factual errors; Grok 4.1 appears to have largely solved that.

In short: Grok just went from “smart and funny” to “the AI you actually want to talk to all day.” If future versions keep this trajectory, the gap in subjective user experience against Claude, Gemini, and GPT could become massive.

Go try it now — it’s live for everyone.