Grok 4.1
xAI's Grok 4.1, tuned for personality, emotional tone and fewer factual errors
Grok 4.1 is a large language model from xAI, released in November 2025 as an update to Grok 4. The company says it keeps the earlier model's reasoning strength while improving creative, emotional and collaborative conversation, and reducing factual errors on information-seeking prompts. It runs in both reasoning and non-reasoning modes.
- Silent rollout ran 1–14 November 2025 across grok.com, X and the mobile apps
- Preferred 64.78% of the time over the previous production model in blind tests
- Thinking variant (quasarflux) reached 1483 Elo on LMArena's Text Arena
- Non-thinking variant (tensor) reached 1465 Elo with no thinking tokens
- Post-training targeted hallucination reduction on information-seeking prompts
What it is
Grok 4.1 is a general-purpose language model built by xAI, the company behind the Grok line. It appeared in November 2025 as the successor to Grok 4. The company describes it as aimed at everyday usability: conversation that feels natural, emotionally aware and consistent in character, while keeping the reasoning ability of earlier versions. It ships in two modes. One, code-named quasarflux, spends thinking tokens before answering. The other, code-named tensor, replies immediately without a thinking stage. Both are offered to the public through grok.com, the X platform and the mobile apps.
How it came about
xAI ran a quiet two-week test of early Grok 4.1 builds from 1 to 14 November 2025, sending them to a growing share of live traffic on grok.com, X and the mobile apps. During that period the team ran blind side-by-side comparisons against the model then serving production. Users preferred the new build 64.78% of the time. The full release followed, with the model appearing in Auto mode and as a named option in the model picker. It builds on the reinforcement learning setup used for Grok 4.
Inside the model
The company says it reused the large-scale reinforcement learning system built for Grok 4, this time targeting style, personality, helpfulness and alignment rather than only verifiable tasks. Because those qualities have no single correct answer to check against, xAI developed a method that uses frontier agentic reasoning models as reward models, scoring and revising responses automatically at scale. Post-training also targeted factual errors on information-seeking prompts. The company reports lower hallucination rates on sampled production queries and on FActScore, a public set of 500 biography questions, testing the non-reasoning model with web search enabled. Layer counts, parameter totals and training data are not given.
What it can do
In blind human preference tests on LMArena's Text Arena, the thinking variant reached 1483 Elo and the non-thinking variant 1465 Elo; xAI reports these as first and second place overall at the time, with Grok 4 previously ranked 33rd. The company also ran EQ-Bench3, an LLM-judged measure of emotional intelligence across 45 roleplay scenarios, and Creative Writing v3, covering 32 writing prompts over three iterations, using rubric scores and normalised Elo. Both were computed with the benchmark's default sampling, the prescribed judge (Claude Sonnet 3.7) and no system prompt. xAI says it is working with the benchmark authors to have the numbers listed.
How to use it
Grok 4.1 is available to all users on grok.com, on X, and in the iOS and Android apps. It rolls out in Auto mode and can also be chosen by name in the model picker. The weights are not published; access is through xAI's own products.
Safety and testing
The company frames part of its post-training work as reducing factual hallucinations, measuring error rates on a stratified sample of real production information-seeking queries and on the FActScore benchmark. Alignment and helpfulness were optimised through reinforcement learning with model-based reward signals. No separate red-teaming programme or safety incident is described.
Reception and debate
xAI reports that during the silent rollout users preferred Grok 4.1 to the model it was replacing 64.78% of the time in blind comparisons, and that it led LMArena's Text Arena at 1483 Elo in thinking mode.
Why it matters
The release is presented as a step in the Grok line, carrying Grok 4's reasoning forward while shifting attention to conversational quality and factual reliability. The company's approach of using frontier reasoning models as automatic reward models for non-verifiable traits is described as a new method developed for this training run.
Questions people ask about Grok 4.1
What is Grok 4.1?
It is xAI's November 2025 update to the Grok 4 language model. The company says it improves conversational quality, personality consistency and emotional awareness while keeping the earlier model's reasoning ability, and reduces factual errors on information-seeking prompts.
How do I use it?
It is available to all users on grok.com, on X, and in the iOS and Android apps. It appears automatically in Auto mode and can also be selected by name in the model picker. The weights are not released.
What are quasarflux and tensor?
These are code names for the two modes of Grok 4.1. quasarflux is the thinking variant, which spends tokens reasoning before answering. tensor is the non-reasoning variant, which returns an immediate response without thinking tokens.
How does it differ from Grok 4?
xAI says it keeps Grok 4's intelligence and reliability but improves style, personality, helpfulness and alignment, using the same reinforcement learning infrastructure. On LMArena's Text Arena, xAI reports Grok 4.1 at 1483 Elo versus Grok 4's overall rank of 33rd.
How was it trained?
xAI applied its large-scale reinforcement learning setup to non-verifiable qualities such as tone and helpfulness, using frontier agentic reasoning models as reward models to judge and iterate on responses automatically. Parameter counts and training data are not published.
Good reading and sources
There is no English Wikipedia article on Grok 4.1 yet, so the facts on this page come from the sources listed above: x.ai/news/grok-4-1. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.