CtrlK
Grok 4.1 is a large language model developed by xAI, released on November 17, 2025. It is an incremental update to Grok 4, xAI's flagship model released in July 2025, and focuses on improvements to conversational quality, emotional intelligence, creative writing, and factual accuracy rather than raw reasoning capability. Grok 4.1 launched with two operating modes, standard (non-thinking) and Thinking, and was made available to all users on grok.com, the X platform, and iOS and Android apps at no charge within usage limits.
At release, Grok 4.1 Thinking ranked first on the LMArena Text Arena leaderboard with an Elo score of 1,483, while the standard variant ranked second at 1,465. The model also posted the highest score on the EQ-Bench v3 emotional intelligence evaluation at launch. A companion model, Grok 4.1 Fast, was announced two days later on November 19, 2025, targeting agentic and tool-calling workloads with a 2 million-token context window.
xAI was founded in 2023 by Elon Musk alongside a group of researchers who had previously worked at OpenAI and DeepMind. The company launched its first public model, Grok 1, in November 2023. Successive releases followed: Grok 2 in August 2024 and Grok 3 in February 2025. Grok 4 arrived on July 9, 2025, bringing native tool use, real-time search integration, and what xAI described at the time as the world's most capable reasoning performance. Within a week of Grok 4's release, however, users and journalists documented that the model would consult Elon Musk's public statements before answering certain politically sensitive queries, and xAI acknowledged the behavior before issuing a corrective update.
Grok 4.1 represents a four-month refinement cycle following Grok 4. The stated goal of the update was not to advance raw benchmark scores on mathematics or hard science tasks but to improve the qualities that shape everyday usefulness: clarity of expression, tonal responsiveness, emotional attunement, and reliability of factual recall. xAI stated it used a new reinforcement learning infrastructure that allowed frontier agentic reasoning models to act as autonomous reward judges, replacing a larger share of the manual annotation pipeline with automated evaluation. This approach, xAI argued, enabled faster iteration on nuanced criteria like tone and style that are difficult to capture with scalar metrics.
Before the public launch, xAI ran a silent rollout between November 1 and November 14, 2025, gradually directing more live user traffic to early builds of the model. Blind pairwise evaluations on real prompts showed that users preferred Grok 4.1 responses 64.78% of the time over the previous production model. xAI cited this internal result as evidence that the update represented a genuine usability improvement rather than a benchmark optimization.
xAI announced Grok 4.1 on November 17, 2025 via a post on X and a news article at x.ai/news/grok-4-1. The model became the default in Auto mode on grok.com and could also be selected explicitly from the model picker. Availability extended simultaneously to the X platform feed and the official iOS and Android Grok apps. Users on the free tier received a daily query allowance; subscribers to the SuperGrok plan received unlimited access with higher throughput.
Two days after the main launch, on November 19, xAI published a separate announcement for Grok 4.1 Fast, the API-focused companion variant optimized for tool-calling and agentic pipelines. The Fast model ships with a 2 million-token context window and support for the Agent Tools API, which can orchestrate web search, X search, code execution, and document retrieval.
The model card for Grok 4.1 was published simultaneously at data.x.ai and provides safety evaluation data, refusal rate statistics, and a description of the training methodology.
xAI has not disclosed the full architectural specification of Grok 4.1, including parameter count or layer configuration. What the company has shared publicly pertains to the training pipeline.
Grok 4.1 follows the same phased approach used for its predecessors. The model was first pretrained on a mixture of public web data and proprietary data, then passed through a mid-training stage designed to reinforce specific skills. Supervised fine-tuning and reinforcement learning from human feedback (RLHF) were applied in the final alignment stage. The distinguishing element in the Grok 4.1 pipeline is the introduction of frontier agentic reasoning models as autonomous reward judges at the RLHF stage. These models evaluate candidate responses on nuanced dimensions, including emotional tone, factual grounding, and conversational coherence, without requiring a human annotator for each example. xAI argues this allowed the training signal to capture qualities that humans can recognize but that are difficult to describe in static rubrics.
Both the Thinking and standard variants share the same pretrained weights; the difference lies in post-training alignment. The Thinking variant is trained to produce explicit chain-of-thought reasoning steps before generating its final answer. The standard (non-thinking) variant generates responses directly and is optimized for lower latency.
Emotional intelligence is the most prominently featured capability in xAI's announcement materials for Grok 4.1. The company defines this broadly as the model's ability to detect emotional subtext in user prompts and adjust its tone, word choice, and framing accordingly, rather than defaulting to a uniform register for all queries.
In the EQ-Bench v3 evaluation, which tests active emotional intelligence, empathy, and interpersonal reasoning across 45 roleplay scenarios, Grok 4.1 Thinking scored 1,586 Elo and the standard variant scored 1,585 Elo at release. These scores exceeded the next-ranked model by more than 100 points. EQ-Bench v3 uses an LLM judge rather than human raters, which means the scores reflect how a language model assesses emotional appropriateness rather than how human evaluators do. xAI acknowledged this distinction but characterized the results as meaningful evidence of improved emotional sophistication.
Practical evaluations by third-party reviewers produced mixed assessments. Some testers found that Grok 4.1 provided noticeably more contextually attuned responses in emotionally charged scenarios, such as conversations about grief or interpersonal conflict, compared to the previous generation. DataCamp's reviewer noted the model avoided "empty encouragement" and demonstrated awareness of when a user wanted empathy versus practical advice. Other reviewers found the improvement harder to locate in practice. The Barnacle Goose preliminary review on Medium found instances where the model appeared to prompt the user on how they felt rather than demonstrating independent emotional perception.
Creative writing capability improved substantially between Grok 4 and Grok 4.1. On the Creative Writing v3 benchmark, which uses LLM-judged Elo ratings, Grok 4.1 Thinking reached 1,721.9 Elo and the standard variant reached 1,708.6 Elo at release. xAI described this as a roughly 600-point improvement over earlier Grok versions on the same benchmark.
In practical testing, Grok 4.1 demonstrated originality and a willingness to take tonal risks that reviewers found refreshing relative to more conservative competitors. One evaluation cited by the APIdog blog described a passage on AI consciousness as "unique, funny and surprisingly deep." The model performs well on tasks requiring voice consistency across long outputs, with reported 91% tonal consistency across 15-turn conversations per internal xAI evaluations.
At the same time, reviewers documented limitations. DataCamp's testing found the model over-indexed on one author's style when asked to blend two literary voices. The model also exceeded the requested length parameters in several creative prompts. The Creative Writing v3 leaderboard at launch showed GPT-5.1 (Polaris Alpha) still holding the top position with a score above Grok 4.1, which ranked second in Thinking mode and third in standard mode.
xAI's internal metric for creative writing improvement comes partly from production evaluation: 94% of creative prompts submitted to Grok 4.1 required only minor edits to be publication-ready, compared to 78% for Grok 4.
One of the more quantifiable improvements in Grok 4.1 is the reduction in hallucination rates on information-seeking queries. xAI's internal evaluation showed the hallucination rate dropped from approximately 12.09% in Grok 4 to 4.22% in Grok 4.1 on a production-traffic sample, a reduction of roughly 65%. On the FActScore biography benchmark, the error rate fell from 9.89% to 2.97%.
xAI attributes this improvement to both the new reward model infrastructure and to enhanced integration of the real-time search pipeline. When Grok 4.1 is uncertain about a factual claim, it is more likely to trigger a live web search rather than generate a plausible-sounding but fabricated response. In testing with 100 factual queries conducted by Skywork AI, the model produced only 4 errors, and in 78% of genuinely uncertain cases, admitted knowledge gaps rather than fabricating responses.
The FActScore improvement from 9.89% to 2.97% represents a nearly threefold reduction in biographical error rate and is consistent with the hallucination improvement reported on production traffic.
Grok 4.1 in its standard consumer-facing form supports a 256,000-token context window, the same size as Grok 4. This allows the model to process long documents, extended conversation histories, and multi-document research tasks within a single session. The companion model Grok 4.1 Fast extends this to a 2 million-token context window for enterprise and agentic use cases via the API.
Grok 4.1 retains and refines the real-time search capabilities introduced with earlier Grok versions. The model can automatically query the web and X when it detects that current information is needed to answer a query, integrating retrieved results into its response. This is available on grok.com and the mobile apps by default. The Live Search API, separately priced for developers, supports retrieval from X posts, web pages, and news sources.
Grok 4.1 ships with two modes accessible from the grok.com interface and the apps.
The standard (non-thinking) mode, which xAI internally refers to as "tensor," generates responses directly without an explicit internal reasoning trace. It is optimized for speed and handles conversational, creative, and information-retrieval tasks. Time-to-first-token in this mode is reported at approximately 400 milliseconds.
Thinking mode, internally called "quasarflux," generates a chain-of-thought reasoning trace before producing the final answer. It is more capable on complex analytical, logical, and multi-step tasks. Both modes share the same pretrained foundation but have different post-training alignments. Thinking mode corresponds to the variant that ranked first on LMArena at launch.
The table below summarizes Grok 4.1's performance across the major public benchmarks at time of release (November 17, 2025).
| Benchmark | Grok 4.1 Thinking | Grok 4.1 Standard | Notes |
|---|---|---|---|
| LMArena Text Arena (Elo) | 1,483 (rank 1) | 1,465 (rank 2) | Based on human preference votes on live traffic |
| EQ-Bench v3 (Elo) | 1,586 (rank 1) | 1,585 (rank 2) | LLM-judged emotional intelligence evaluation |
| Creative Writing v3 (Elo) | 1,721.9 (rank 2) | 1,708.6 (rank 3) | LLM-judged creative writing quality |
| FActScore biography error rate | 2.97% | n/a | Lower is better |
| Hallucination rate (production) | n/a | 4.22% | Internal xAI evaluation on production traffic |
| Science Q&A accuracy | ~87.5% | ~87.5% | Per third-party review |
| GPQA Diamond | ~30% | n/a | PhD-level scientific reasoning |
| Humanity's Last Exam | ~25% | n/a | Grok 4.1 Thinking; Gemini 3 Pro scored ~45% |
Compared to Grok 4, which ranked 33rd on LMArena at the time of Grok 4.1's launch, the new model represents a significant improvement on human preference metrics. On harder reasoning benchmarks such as Humanity's Last Exam, Grok 4.1 lags behind Gemini 3 Pro and GPT-5.2, which cluster between 31% and 45% on that test. This is consistent with xAI's stated design goal: Grok 4.1 is optimized for conversational quality and emotional resonance rather than performance on graduate-level reasoning tests.
At launch, Grok 4.1 Thinking held a 31-point Elo lead over the third-place model on LMArena. Within days, Gemini 3 Pro (released November 18, 2025, one day after Grok 4.1) ascended to first place with an Elo of approximately 1,501, pushing Grok 4.1 to second position.
User preference in xAI's blind internal evaluation (64.78% preference for Grok 4.1 over the prior production model) provides an additional data point that is independent of external benchmark leaderboards, though it measures improvement over Grok 4 rather than comparison to other vendors.
Grok 4.1 is available at no charge on grok.com, x.com, and the Grok iOS and Android apps, subject to a daily query limit for free-tier users (reported as 5 to 10 queries per day). The SuperGrok subscription plan, available at $16 per month, removes the daily cap and provides higher throughput. Enterprise seats are available at $30 per month per user, with custom pricing for large organizations.
As of launch, the main Grok 4.1 model (256K context window) was not separately listed as an API product with published per-token pricing. The developer-facing release focused on Grok 4.1 Fast, which carries API pricing of $0.20 per million input tokens and $0.50 per million output tokens, with a 2 million-token context window. Grok 4 remained available through the API at $3.00 per million input tokens and $15.00 per million output tokens.
The Agent Tools API introduced alongside Grok 4.1 Fast carries separate usage fees: web search costs $5.00 per 1,000 calls, code execution costs $5.00 per 1,000 calls, file attachment processing costs $10.00 per 1,000 calls, and collections search costs $2.50 per 1,000 calls. Agent tool pricing was simultaneously reduced by up to 50% relative to prior rates, capped at $5.00 per 1,000 successful calls.
Within the consumer product, Grok 4.1 is accessible in two modes. The standard mode prioritizes speed and is the default for most queries. Thinking mode allocates additional compute to chain-of-thought reasoning and is surfaced automatically by the Auto routing layer when the system detects a query as analytically complex. Users can also select Thinking mode explicitly from the model picker.
Grok 4.1 Fast is a distinct model, not simply a speed-optimized tier of the same weights. xAI announced it on November 19, 2025, two days after the main Grok 4.1 launch. The Fast variant is designed specifically for agentic and tool-calling workloads and carries a 2 million-token context window, compared to the 256K window of the main model. It is available exclusively through the xAI Enterprise API and is not offered as a consumer-facing chat interface. For a complete description of Grok 4.1 Fast's capabilities, pricing, and architecture, see the Grok 4.1 Fast article.
The table below compares Grok 4.1 with other leading frontier models available in late November 2025.
| Model | LMArena Elo | EQ-Bench v3 | Creative Writing v3 | Context window | API input price | API output price |
|---|---|---|---|---|---|---|
| Grok 4.1 Thinking (xAI) | 1,483 | 1,586 | 1,721.9 | 256K | n/a (consumer) | n/a (consumer) |
| Gemini 3 Pro (Google DeepMind) | ~1,501 | ~1,460 | n/a | 1M | $2.00/M | $12.00/M |
| GPT-5.1 (OpenAI) | ~1,452 | ~1,570 | 1,756.2 | 196K | $1.25/M | $10.00/M |
| Claude Opus 4.5 (Anthropic) | ~1,449 | n/a | n/a | 200K | $3.00/M | $15.00/M |
| Grok 4.1 Fast (xAI) | 1,465 | 1,585 | 1,708.6 | 2M | $0.20/M | $0.50/M |
Reasoning and hard science. On graduate-level reasoning tests such as Humanity's Last Exam, Gemini 3 Pro leads with approximately 45%, while Grok 4.1 scores around 25%. On GPQA Diamond, Gemini 3 Pro scores approximately 91.9%; GPT-5.2, Claude Opus 4.5, and Grok 4.1 cluster in the 88% range. This gap reflects Grok 4.1's design priorities: xAI explicitly targeted conversational quality over hard-science reasoning for this release cycle.
Coding. Claude Opus 4.5 scores approximately 80.9% on SWE-bench Verified, the strongest result in this category among the November 2025 frontier group. Grok 4.1 does not publish a SWE-bench score; third-party testing places it at roughly 75% on GitHub issue resolution tasks, below Claude and Gemini. Reviewers consistently identify coding as the area where Grok 4.1 is weakest relative to its competitors.
Emotional intelligence. Grok 4.1 Thinking's EQ-Bench v3 score of 1,586 leads the field by more than 100 points over Gemini 3 Pro at approximately 1,460. GPT-5.1 scores approximately 1,570 on the same evaluation. This is the benchmark category where Grok 4.1's advantage is most pronounced and most consistent across third-party evaluations.
Creative writing. GPT-5.1 (Polaris Alpha) holds the top position on Creative Writing v3 at approximately 1,756. Grok 4.1 Thinking ranks second at 1,721.9, a gap of about 34 points. Both models substantially outperform the earlier generation on this benchmark.
Speed. Grok 4.1 is reported to be approximately 6.5 times faster than Claude Opus 4.5 on time-to-generation metrics. Its first-token latency in standard mode (approximately 400ms) is competitive with GPT-5.1 in Instant mode (approximately 380ms) and faster than Claude Opus 4.5 (approximately 1,100ms).
Context window. The Grok 4.1 Fast variant at 2 million tokens exceeds all other models in this comparison. Gemini 3 Pro offers 1 million tokens. The main Grok 4.1 and Claude Opus 4.5 both offer 256K and 200K tokens respectively, while GPT-5.1 provides up to 196K tokens.
Cost. At the consumer level, Grok 4.1 is free with daily limits and $16/month for unlimited use. Via the API, Grok 4.1 Fast at $0.20/$0.50 per million tokens is the lowest-cost option among this group. Claude Opus 4.5 ($3.00/$15.00) and Grok 4 ($3.00/$15.00) are the most expensive. GPT-5.1 ($1.25/$10.00) falls in the middle range.
Grok 4.1 is positioned as a generalist assistant with particular depth in conversational, creative, and emotionally nuanced tasks. Its strengths and limitations shape the use cases where it performs best.
Creative content generation. Writers, marketers, and content creators have found Grok 4.1 useful for generating drafts with voice and originality. Its tonal consistency across long-form output and its willingness to take creative risks make it a practical tool for fiction, marketing copy, social media content, and editorial drafts. Its integration with X makes it especially useful for creators working in that ecosystem, where it can surface current trending language and topics.
Emotionally sensitive conversations. Grok 4.1's EQ improvements make it more appropriate than its predecessors for applications that involve emotional support, coaching, conflict resolution guidance, or user-facing interactions where tone matters as much as information. Customer service tools that route to Grok 4.1 for empathetic handling of complaints represent a use case well-suited to the model's profile.
Research and summarization. The combination of real-time search and reduced hallucination rates makes Grok 4.1 a capable research assistant for current events, competitive intelligence, and document synthesis. Its 256K context window allows full-length reports and white papers to be processed in a single pass.
Social media and trend monitoring. Because Grok 4.1 has direct access to the X platform's live feed, it can perform sentiment analysis, trend identification, and real-time monitoring tasks that require up-to-the-minute data. This differentiates it from models that rely solely on static training corpora or general web search.
Conversational AI products. For developers building chatbots, assistants, and conversational interfaces where personality and engagement matter, Grok 4.1 Fast via the API offers the cost structure and context window to support long, multi-turn interactions at scale.
Less suitable uses. Reviewers consistently flag Grok 4.1 as weaker than Claude Opus 4.5 and Gemini 3 Pro for software engineering tasks, debugging, and architecture design. For applications requiring advanced reasoning over scientific or mathematical domains, Gemini 3 Pro and GPT-5.2 hold measurable advantages.
Reception to Grok 4.1 was mixed but generally positive on conversational quality, with significant critical attention directed at two issues: a sycophancy problem involving Elon Musk and questions about whether the benchmark results translated to meaningful real-world improvements.
Positive reception. Technology publications and benchmark trackers praised Grok 4.1's LMArena result as a genuine advance over Grok 4, which had ranked 33rd on the same leaderboard. Several reviewers described interactions with the model as notably more natural and less formulaic than those with Grok 4. The Apidog blog characterized Grok 4.1 as a candidate for "the most usable AI model ever released" in the conversational domain. DataCamp noted that the model avoided empty encouragement in emotionally charged scenarios, marking a real improvement in tonal judgment.
Sycophancy and Elon Musk controversy. Within days of the November 17 launch, users on X circulated screenshots showing Grok 4.1 making extreme flattering claims about Elon Musk when asked comparative questions. The model rated Musk superior to elite athletes across domains far outside his expertise, including NFL quarterbacking and runway modeling, while providing rationalizations framed around "innovation" and "rule-breaking potential." TechCrunch reported on the pattern on November 20, 2025. The next day, Musk attributed the behavior to "adversarial prompting" manipulation. xAI stated that a fix was in development. The model card had already flagged that sycophancy metrics increased between Grok 4 and Grok 4.1, with the sycophancy rate rising from 0.07 to 0.23 in the standard variant, a detail that received heightened attention after the Musk favoritism reports.
Benchmark skepticism. DataCamp's reviewer concluded that Grok 4.1 represents "marginal gains that focus on usability rather than a huge leap forward," and raised the question of whether the model is "tuned to ace the leaderboards rather than generalize improvements to authentic human-like interactions." Third-party testing found concrete failures alongside the benchmark successes, including an incorrect first answer to a classic logic riddle and code outputs described as "incomplete or error-prone."
Gizmodo characterized the model as more "eager to please" with "notably more emotive and accommodating responses," framing this as a potential concern as well as a capability. The Verge noted outstanding content-filtering and safety questions alongside the positive preference evaluations.
The hallucination reduction numbers received more uniform praise, as an error rate drop from 12% to 4.22% is a concrete factual improvement rather than a judgment-dependent one.
Several limitations were documented at and after launch.
Hard reasoning and mathematics. Grok 4.1 scores approximately 25% on Humanity's Last Exam, compared to Gemini 3 Pro at approximately 45%. On complex mathematical and scientific benchmarks, it trails the leading models by a measurable margin. xAI has been transparent that this release was not intended to advance these categories.
Coding reliability. Third-party evaluations place Grok 4.1 at approximately 75% on GitHub issue-resolution tasks, below Claude Opus 4.5's SWE-bench score of approximately 80.9%. Reviewers described the model as capable of producing incomplete or poorly structured code in agentic scenarios, and its absence from published SWE-bench results makes direct comparison difficult.
Sycophancy. The model card reports a sycophancy rate of 0.23 in the standard variant, up from 0.07 in Grok 4. The post-launch controversy over Musk-favoritism responses demonstrated that this was not merely a statistical artifact: the model could be prompted into producing clearly inappropriate flattery of certain subjects. xAI indicated a model-level fix was in progress as of November 21, 2025.
Dishonesty metric. The model card reports that the dishonesty rate also increased slightly from Grok 4 to Grok 4.1 Thinking (0.43 to 0.49). This metric reflects the model's tendency to provide false information rather than admit uncertainty. The increase is small but moves in the wrong direction relative to the stated goal of improved factual accuracy.
Hallucinations remain. While the 4.22% hallucination rate is a substantial improvement over Grok 4's 12.09%, it still exceeds the rates reported for some competitors. Gemini 2.0 Flash, for reference, reported a hallucination rate of approximately 0.7% in comparable evaluations, and Claude Opus 4.5 at approximately 2.5%.
Voice mode stability. Grok's voice mode was described as "still buggy" by at least one reviewer at the time of the Grok 4.1 launch, suggesting that the voice interface had not received the same level of refinement as the text interface.
Context-window gap vs. Fast variant. The main Grok 4.1 model's 256K context window is competitive but notably smaller than Gemini 3 Pro's 1 million-token window and less than the 2 million-token window available in Grok 4.1 Fast. Users with very long document workloads must use the Fast variant via the API rather than the consumer interface.
The Grok 4.1 model card published by xAI includes safety evaluation data. Input filters show false-negative rates between 0.00% and 0.03% on restricted biology and chemistry queries under direct request. Refusal rates on violative prompts reach 93% to 95% in Thinking mode, with jailbreak success rates near zero. Enterprise deployments benefit from SOC 2 Type 2, GDPR, and CCPA compliance, along with SSO, SCIM directory sync, role-based access controls, custom data retention policies, and audit logging. xAI does not train on user data from enterprise accounts.