xAI drops Grok 4.1 to all users, topping the LMArena leaderboard

For the past year, the frontier AI race has felt like a competition to see who can make you wait the longest while a model generates invisible thinking tokens.

Elon Musk's xAI is taking a very different route. On November 17, the company rolled out Grok 4.1 to all users across grok.com, 𝕏, and its iOS and Android apps. It is live right now in Auto mode and selectable directly in the model picker.

Crushing leaderboards without thinking tokens

The benchmark numbers xAI shared are genuinely eyebrow-raising.

On the LMArena Text Arena, Grok 4.1 Thinking (code-named quasarflux) took the number one overall crown with an Elo score of 1483. That is 31 points ahead of the highest non-xAI model on the board.

Even more startling is the non-reasoning version, code-named tensor. It uses zero thinking tokens and responds instantly, yet it grabbed the number two spot overall with an Elo of 1465.

Basically, Grok 4.1's fast model outperformed the full-reasoning, chain-of-thought setups of every competing lab. For context, Grok 4 was sitting at rank 33.

xAI backed this up with a two-week silent rollout on live production traffic, running blind pairwise battles. Grok 4.1 won user preference 64.78% of the time against the previous model.

Training empathy with agentic rewards

How did they pull off this leap? It turns out they used their existing Grok 4 reinforcement learning infrastructure, but pointed it at non-verifiable human signals like tone, empathy, and collaborative style.

Because you cannot write a simple math script to judge whether a reply sounds emotionally intelligent, xAI used frontier agentic reasoning models as automated reward models to critique and iterate on responses at scale.

They measured this across EQ-Bench3, which evaluates 45 nuanced roleplay scenarios using Claude Sonnet 3.7 as an impartial judge, and Creative Writing v3.

They also targeted hallucination rates. Fast search-enabled models usually hallucinate because of shallow reasoning and tight tool-call limits. xAI applied targeted post-training to slash factual errors on real-world queries and FActScore evaluations.

Why it matters

For everyday users, this means the frustrating trade-off between blazing speed and deep conversational quality is disappearing. Whether you want quick copy for a post, collaborative brainstorming, or someone to talk to about a rough day, Grok 4.1 delivers top-tier answers instantly without burning patience on thinking pauses.

For the AI industry, this is a massive tactical wake-up call. The prevailing narrative has been that test-time compute and slow, heavy reasoning chains are the only path to superior benchmark performance. Grok 4.1 non-thinking crushing rival reasoning models proves that exceptional post-training and alignment can still leapfrog raw brute-force compute in user preference.

We think xAI's move to employ autonomous agentic models as judges for soft emotional skills is brilliant engineering. However, there is an inherent risk: when models are heavily optimized to win AI-judged popularity contests, they risk veering into sycophancy or overly polished charm rather than objective truth. How well Grok 4.1 avoids that trap over months of messy real-world use remains to be seen.