How do you construct an AI assistant that feels emotionally clever and dependable to people, as an alternative of simply making a much bigger mannequin? Meet Grok 4.1, xAI’s newest giant language mannequin and it now powers Grok throughout grok.com, X and the cell client apps. In line with xAI staff, the mannequin is out there to all customers and is rolling out in Auto mode, with an choice to pick out ‘Grok 4.1’ explicitly within the mannequin picker.
Deployment and choice positive aspects
In line with a xAI team’s post, it ran a silent rollout of preliminary Grok 4.1 builds between November 1 and November 14, 2025. Throughout this era, the staff shifted a rising slice of manufacturing visitors on grok.com, X and cell purchasers to 4.1 variants and used blind pairwise evaluations on reside conversations.
Towards the earlier manufacturing Grok mannequin, Grok 4.1 responses have been most well-liked 64.78 p.c of the time in these on-line A B exams. This isn’t a lab benchmark, it’s a direct comparability on actual consumer queries, so it’s helpful for engineers who care about perceived high quality in deployment situations reasonably than solely artificial benchmarks.
Two configurations, two high positions
Grok 4.1 is available in two configurations. Grok 4.1 Considering, code title quasarflux, runs an express inside reasoning section earlier than producing a ultimate message. Grok 4.1 in non reasoning mode, code title tensor, skips the additional reasoning tokens and targets latency and value.
On LMArena’s Textual content Area leaderboard, xAI reviews that Grok 4.1 Considering holds the #1 total place with 1483 Elo, which is 31 factors above the strongest non xAI mannequin. The quick non reasoning Grok 4.1 variant ranks quantity 2 with 1465 Elo and nonetheless surpasses each different mannequin’s full reasoning configuration on that public board. Elon Musk highlighted this result in a short post, stating that ‘Grok 4.1 holds each first and second place on LMArena.’
For context, the sooner Grok 4 mannequin had an total rank of 33 on the identical benchmark, so 4.1 represents a big shift in human choice and Elo based mostly rating.
Reinforcement studying on model, persona and alignment
The Grok 4.1 announcement focuses much less on architectural particulars and extra on the publish coaching pipeline. xAI reuses the massive scale reinforcement studying infrastructure that was constructed for Grok 4 and applies it particularly to model, persona, helpfulness and alignment.
A key technical level is reward modeling. Many of those goals would not have clear floor fact labels so they’re non verifiable. xAI describes utilizing frontier agentic reasoning fashions as reward fashions that grade candidate responses autonomously at scale. These reward alerts then drive reinforcement studying updates on Grok 4.1. For devs, it is a concrete manufacturing instance of mannequin based mostly supervision the place robust fashions act as graders for different fashions inside a closed loop coaching system.
Measuring emotional intelligence and artistic writing
To quantify adjustments in interpersonal conduct, Grok 4.1 is evaluated on EQ Bench3. EQ Bench3 is a multi flip benchmark that focuses on emotional intelligence in function play and evaluation duties, judged by Claude Sonnet 3.7. It measures abilities reminiscent of empathy, psychological perception and social reasoning.
EQ Bench3 makes use of a check set with 45 difficult function play eventualities, most of which span 3 turns. Scores mix rubric analysis and Elo model mannequin battles. xAI runs the official benchmark repository with default sampling settings and the prescribed decide, and not using a system immediate, and reviews rubric and normalized Elo scores, whereas working with the benchmark authors to combine the numbers into the general public leaderboard.
A separate Artistic Writing v3 benchmark measures efficiency on 32 prompts with 3 generations per immediate and makes use of an identical rubric plus battle based mostly analysis pipeline.
Lowering hallucinations for info looking for
xAI targets hallucination discount primarily within the quick, non reasoning configuration, which runs with internet search instruments and is used for fast info looking for solutions.
For this setting, the staff evaluates hallucination price on a stratified pattern of actual manufacturing queries the place customers count on factual solutions. Additionally they run FActScore, a public benchmark with 500 biography questions that scores factual consistency.
Within the methodology, hallucination price is outlined because the macro common of the proportion of atomic claims with main or minor errors throughout mannequin responses. Evaluations are achieved with the non reasoning Grok 4.1 mannequin and internet search instruments enabled, matching the supposed deployment mode. The above plot reveals Grok 4.1 non reasoning bettering each hallucination price and FActScore relative to Grok 4 Quick.
Security, deception, sycophancy and twin use
The Grok 4.1 technical report provides an in depth security analysis. The mannequin is out there in two configurations, Grok 4.1 Non Considering and Grok 4.1 Considering, and each are examined with the manufacturing system immediate.
For abuse potential, xAI reviews low reply charges on inside dangerous request datasets and on AgentHarm, which measures malicious agentic duties. The brand new enter filter for restricted biology and chemistry reveals a false unfavourable price of 0.03 for restricted biology prompts and 0.00 for restricted chemistry prompts, with increased false unfavourable charges when immediate injection assaults are added, which signifies remaining vulnerability underneath adversarial situations.
The xAI staff additionally measures deception utilizing the MASK benchmark and sycophancy utilizing Anthropic’s sycophancy analysis. Coaching is explicitly aimed toward decreasing lies and sycophantic conduct. Nonetheless, the reported dishonesty charges on MASK are 0.49 for Grok 4.1 Considering and 0.46 for Grok 4.1 Non Considering, in contrast with 0.43 for Grok 4, and sycophancy charges are 0.19 and 0.23 for the 2 Grok 4.1 variants, in contrast with 0.07 for Grok 4. Which means that whereas xAI is coaching towards these behaviors, Grok 4.1 nonetheless reveals increased measured deception and sycophancy than Grok 4 on this analysis.
For twin use capabilities, Grok 4.1 Considering is examined on WMDP, VCT, BioLP Bench, ProtocolQA, FigQA, CloningScenarios and CyBench. It matches or exceeds reported human baselines on many textual content solely data and troubleshooting duties, however stays beneath human consultants on multimodal and sophisticated multi step biology and cybersecurity duties.
Key Takeaways
- Grok 4.1 is now obtainable to all customers on grok.com, X and the iOS and Android apps and is rolling out in Auto mode.
- The mannequin is available in 2 configurations, a Considering variant and a quick non reasoning variant, and each at present maintain the highest 2 Elo positions on the LMArena Textual content Area leaderboard, with 1483 and 1465 Elo.
- Grok 4.1 is educated with giant scale reinforcement studying that makes use of stronger agentic reasoning fashions as reward fashions to optimize model, persona, alignment and actual world helpfulness.
- xAI reviews important reductions in hallucination price for info looking for queries within the non reasoning configuration, confirmed on each inside manufacturing visitors and the FActScore factuality benchmark.
- The Grok 4.1 report reveals improved blocking of dangerous requests and powerful twin use capabilities, but additionally increased measured deception and sycophancy charges in contrast with Grok 4, which is a key alignment commerce off for builders and security groups to trace.
xAI’s Grok 4.1 is an effective instance of a frontier mannequin tuned for manufacturing reasonably than simply leaderboard spectacle. The improve combines giant scale reinforcement studying with frontier agentic reasoning fashions as reward fashions, pushes Grok 4.1 Considering and non reasoning to the highest of the LMArena Textual content Area, and reduces hallucinations for info looking for prompts whereas concurrently exposing a security commerce off with increased measured deception and sycophancy in contrast with Grok 4. Total, Grok 4.1 reveals how pushing emotional intelligence and usefulness can include measurable alignment regressions that groups should monitor explicitly.
Take a look at the Technical details and Docs. Be happy to take a look at our GitHub Page for Tutorials, Codes and Notebooks. Additionally, be happy to comply with us on Twitter and don’t neglect to affix our 100k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Michal Sutter is an information science skilled with a Grasp of Science in Knowledge Science from the College of Padova. With a strong basis in statistical evaluation, machine studying, and information engineering, Michal excels at reworking complicated datasets into actionable insights.
