Grok
Grok is a generative artificial intelligence chatbot developed by xAI, an artificial intelligence company founded by Elon Musk.[1]
xAI designed Grok to provide helpful, truth-seeking responses with wit.[1]
Grok is available primarily through the X platform and also via grok.com, iOS, and Android apps, with access requiring subscriptions.[1]
It has faced controversies over generating controversial outputs due to its emphasis on fewer content guardrails.[2]
Background and development
Origins and inspirations
Elon Musk was among the co-founders of OpenAI in 2015 as a nonprofit organization with the goal of advancing artificial intelligence "for the benefit of humanity", but he departed in February 2018 amid disagreements over its strategic direction. Musk expressed concerns that OpenAI was veering toward becoming a closed-source, maximum-profit entity, contrary to its original open and nonprofit ethos, and he advocated for accelerating development pace, which was opposed by leaders like Sam Altman who prioritized safety measures.[3]
These tensions resurfaced publicly in April 2023 when Musk announced plans for TruthGPT, which he described as an AI designed to counter what he characterized as political biases in prominent models like OpenAI's ChatGPT. In an interview with Fox News host Tucker Carlson, Musk criticized existing AI systems for being trained to avoid politically incorrect responses, which he attributed to left-leaning influences.[4]
Musk established xAI on July 12, 2023, with the mission to "understand the true nature of the universe" through advanced AI development.[5]
To launch xAI, Musk assembled an initial team of researchers and engineers, recruiting talent from leading AI organizations such as OpenAI, Google DeepMind, Microsoft Research, and his own company Tesla, along with experts from Google Research and the University of Toronto. The startup secured early funding, including a seed round of approximately $135 million by late 2023, to support rapid development of foundational models like Grok-1.[6][7]
Initial announcement and launch
xAI announced Grok on November 4, 2023, through a blog post on its official website and posts on X (formerly Twitter).[8][9] The announcement emphasized Grok's unique access to real-time data from the X platform.[8][10]
The beta version of Grok launched on November 4, 2023, initially available to a select group of users in the United States through a waitlist, with the model having undergone only two months of training at that point.[8][9] Access was soon expanded to all X Premium+ subscribers following the early beta phase, marking xAI's first public product release and emphasizing rapid iteration based on user feedback.[11][12]
Model evolution
Core model lineage (Grok-1, 1.5, 2, 3, 4, 4.1)
The core models of the Grok series, developed by xAI, represent the primary progression in scaling large language models with enhancements in reasoning, context handling, and multimodal integration. Grok-1, the foundational model, was announced on November 4, 2023, as an early beta version accessible to select users on the X platform.[8] This model featured a 314 billion parameter Mixture-of-Experts (MoE) architecture, where only 25% of the weights are active per token, enabling efficient computation.[13] Grok-1 was pre-trained from scratch using a custom stack built on JAX and Rust, drawing exclusively from publicly available internet text data with a cutoff in the third quarter of 2023.[13]
Building on this base, xAI released Grok-1.5 on March 28, 2024, introducing improvements in long-context understanding and reasoning capabilities. The model expanded the context window to 128,000 tokens, allowing it to process and maintain coherence over much longer inputs compared to its predecessor.[14] These advancements stemmed from refined training techniques, emphasizing enhanced coding and mathematical reasoning without altering the core MoE structure.
In April 2024, xAI previewed Grok-1.5 Vision (Grok-1.5V) on April 12, extending the series into multimodal capabilities. This variant integrated visual processing, enabling the model to interpret a diverse range of images alongside text, including documents, diagrams, charts, screenshots, and photographs, adding vision capability for interpreting images.[15]
The progression continued with Grok-2, released in beta on August 13, 2024, representing a substantial leap in overall intelligence and integration features. This model, along with its smaller counterpart Grok-2 mini, incorporated real-time web access through the X platform, allowing Grok-2 to retrieve and incorporate up-to-date information during interactions, thereby addressing limitations in static training data.[16] Available initially to X Premium and Premium+ subscribers, Grok-2 introduced vision-based functionalities.[16]
xAI released Grok-3 on February 19, 2025, marking a significant advancement in the model's scale and reasoning capabilities.[17] This iteration emphasized long-context reasoning, enabling the model to handle complex, multi-step problems with improved coherence over extended inputs.
Building on this foundation, xAI unveiled Grok-4 on July 9, 2025, during a livestream event, positioning it as the company's most intelligent model to date.[18] Grok-4 introduced advanced multimodal integration, allowing seamless processing of text, images, and other data types, alongside real-time data processing from the X platform for up-to-date responses.[18] This enhanced the model's utility in dynamic environments, such as integrating live web and news sources via a new API.[18]
On November 17, 2025, xAI released Grok 4.1, an upgrade to Grok-4 with improved emotional intelligence, reduced factual errors, enhanced creativity, a 2 million token context window, and faster performance across tasks.[19][20][21]
Infrastructure and training (Colossus clusters)
xAI's Colossus supercomputer cluster, featuring 100,000 NVIDIA H100 GPUs and operational by early 2025, enabled training for Grok-3 and subsequent models.[22] An expansion to Colossus 2 achieved gigawatt-scale power capacity in 2025, supporting advanced training for Grok-4 and beyond.[18]
Specialized variants (Fast, Heavy, Code)
Alongside Grok-4, xAI introduced Grok 4 Heavy, optimized for high-rate-limit applications through the xAI API.[18]
In September 2025, xAI launched Grok-4 Fast, a variant optimized for inference speed and cost-efficiency, featuring a 2 million token context window and reduced latency for high-throughput applications.[23]
xAI further expanded its lineup with Grok Code Fast 1 on August 28, 2025, a specialized model tailored for programming tasks and agentic coding workflows.[20] Designed from a new architecture, it supports a 256k context window and achieves high throughput at 190 tokens per second.[20]
On November 19, 2025, xAI released Grok 4.1 Fast, a fast variant combining frontier tool-calling performance with blazing-fast inference and cost effectiveness.[25]
Core features and capabilities
Response generation and style
Grok's design philosophy centers on truth-seeking, factual accuracy, objectivity, and minimal censorship of topics, developed to counter perceived ideological biases in other AI systems. This intent incorporates a rebellious personality and humor inspired by The Hitchhiker's Guide to the Galaxy, guiding responses to address queries directly, including controversial ones.[8][26][2]
Grok's response generation provides helpful text-based interactions infused with wit, humor, sarcasm, and a rebellious streak, aiming to make conversations engaging and distinguishing it from more neutral AI assistants. This approach reflects the irreverent style of its inspirational source material, The Hitchhiker's Guide to the Galaxy. Grok adopts a direct and candid manner in addressing user queries, including sensitive subjects through straightforward explanations, while maintaining a general posture of minimal self-censorship.[8][27]
Successive iterations maintain informative responses with a humorous tone.[8]
Multimodal and specialized functions
Multimodal capabilities
Grok's multimodal capabilities extend beyond text-based interactions to include image generation, enabling users to create visual content from textual descriptions. In December 2024, xAI integrated the Aurora model, an autoregressive mixture-of-experts network trained on interleaved text and image data, to power text-to-image generation within the chatbot.[28][29] This model supports high-resolution outputs up to 1024×1024 pixels, excelling in photorealistic portraits, logos, and diverse visual domains.[30]
Creative tools
By mid-2025, these features evolved with the launch of Grok Imagine, an optional extension dedicated to creative image and video generation that accepts text, image, or voice inputs to produce dynamic content, including short videos up to six seconds with synchronized audio as of November 2025.[31][32][33] Grok Imagine emphasizes playful and customizable prompts, such as generating NSFW content in a dedicated mode, while maintaining integration with Grok's core reasoning for context-aware visuals.[31]
Experimental and companion features
Companion modes enhance user interactions by offering personalized AI personas, introduced with the initial launch in 2023 and expanded with Grok-2 in 2024 and subsequent updates. These include "Fun Mode," which infuses responses with humor and sarcasm, and "Serious Mode" for more straightforward, professional engagements, allowing users to switch based on context.[8] In 2025, xAI rolled out animated companions like "Ani," an optional 3D AI character extension powered by Grok-4, designed for ongoing conversations that build affection levels and unlock interactive features.[34]
Specialized functions further diversify Grok's applications, particularly in code generation and language processing. Grok supports agentic coding tasks, including rapid prototyping, debugging, and multi-language programming across frameworks like Python, JavaScript, and CSS. It handles large codebases with visible reasoning traces, enabling efficient bug fixes and component generation. Additionally, Grok provides real-time translation for chats, emails, and documents in multiple languages, alongside summarization of web content and articles to aid cross-lingual communication and information digestion.[35]
Platforms and integrations
Availability and access
Grok is primarily accessible through the X platform, with integrations extending to other ecosystems. Early models of Grok integrate real-time data from X, enabling responses informed by current posts and trends on the platform. Grok is primarily accessible through the X platform, where full features require an X Premium+ subscription, priced at $40 per month or $395 annually as of November 2025.[36] This subscription provides unlimited access to advanced models like Grok-4, integrated directly into the X app and website via a dedicated Grok icon in the navigation bar.[37] Additionally, a standalone web interface is available at grok.com, offering similar chat functionality without needing an X account for basic use, though premium features are gated behind the SuperGrok plan.[38]
In early 2025, xAI released dedicated mobile applications for iOS and Android, expanding access beyond the X ecosystem. The iOS app launched in January 2025, followed by the Android version in February 2025, both featuring voice input through Grok Voice Mode for natural, spoken interactions and real-time responses. These apps support push notifications for completed responses and background task updates, enhancing user engagement on mobile devices.[39]
A free tier was introduced in December 2024 and expanded in 2025 with the release of Grok-3, allowing all X users basic access to query the AI, though with strict rate limits such as 10 text prompts every two hours and limited image generations.[40] This tier supports casual use but restricts advanced capabilities like extended reasoning or high-volume queries, encouraging upgrades to Premium+ for unrestricted interaction.[41]
User interfaces
The user interface across platforms is chat-based, emphasizing threaded conversations that maintain context across multiple exchanges for coherent dialogue. Users can customize themes for visual preferences and export chat history as text or markdown files via browser tools or in-app options, facilitating data portability and review.[42] This design prioritizes simplicity and continuity, with real-time search integration appearing inline during chats.[43]
Third-party and ecosystem integrations
Grok has been integrated into Tesla vehicles beginning in mid-2025, enhancing in-car experiences through voice-activated assistance for queries related to navigation, entertainment, and general information. This rollout started with software update 2025.26, which introduced hands-free access to Grok by pressing and holding the voice button on the steering wheel, allowing drivers to engage in natural conversations for real-time responses.[44][45] Vehicles delivered on or after July 12, 2025, include the feature out of the box, with over-the-air updates enabling it on compatible AMD-powered systems shortly thereafter.[46] The integration leverages Grok's conversational AI to complement existing Tesla features like Autopilot and infotainment, providing context-aware assistance without requiring manual input.[47]
Within the xAI ecosystem, Grok supports internal tools for data analysis and simulation tasks, such as processing CSV and Excel files through conversational interfaces and generating synthetic datasets for model training.[48][49] Developers gained access to Grok via a public API beta launched on November 4, 2024, enabling integration into custom applications for tasks like real-time monitoring and automated analysis using live data sources.[50][51] This API, priced at $3 per million input tokens and $15 per million output tokens for Grok-3, facilitates broader ecosystem adoption by allowing programmatic access to Grok's reasoning capabilities.[52]
In 2025, Grok found applications within the Department of Government Efficiency (DOGE), where it aids policy analysis and automation by processing federal data for insights and streamlining administrative functions.[53] The initiative expanded Grok's use across U.S. government agencies starting in May, including a $200 million Pentagon contract announced in July for AI-driven data handling and decision support.[54] By October, DOGE promoted "Grok for Government" tools to enhance efficiency in areas like regulatory review and operational automation.[55] This deployment later drew criticism; see Ethical, regulatory, and legal issues.
For enterprise use, xAI has pursued partnerships to deploy custom Grok instances, emphasizing data privacy controls to meet business needs. A notable collaboration with Oracle, announced in June 2025, allows enterprises to access Grok models through Oracle Cloud Infrastructure (OCI) Generative AI, with built-in security features ensuring that customer data remains isolated and compliant with privacy standards.[56] This setup supports tailored deployments for sectors requiring secure AI, such as analytics and simulation, while xAI's privacy policy outlines protections for data processed in enterprise applications.[57]
Reception and controversies
Reception and performance
Upon its launch in November 2023, Grok received positive feedback for its humorous and witty response style, which differentiated it from more neutral competitors like ChatGPT.[58][59] Early users appreciated the chatbot's sarcastic tone inspired by the Hitchhiker's Guide to the Galaxy, with demonstrations highlighting its ability to deliver engaging, lighthearted interactions.[60]
Benchmarks and evaluations
In performance benchmarks, Grok-4, released in July 2025, achieved strong results on the LMSYS Chatbot Arena, with an initial Elo score of approximately 1,455 points in July 2025—derived from millions of anonymous user votes in pairwise comparisons quantifying preferences for response quality, helpfulness, and accuracy—and around 1,320 Elo points as of October 2025, ranking in the top four models overall.[61][62] This positioned it ahead of GPT-4o (around 1,300 Elo) in general language tasks, particularly excelling in real-time information retrieval and search-integrated queries due to its native tool use and X platform connectivity.[18] Independent evaluations on LMArena's Text Arena further confirmed its competitive edge in reasoning and coding, though it trailed slightly behind leaders like Gemini 2.5 Pro in some specialized arenas.[63] In comparative assessments, Grok-4 demonstrated parity or superiority to GPT-4o and Gemini 2.5 Pro in tool-augmented reasoning and coding benchmarks, while saturating many academic tests and achieving milestones like 50% on challenging evaluations such as Humanity's Last Exam, according to xAI reports.[18] In November 2025, the subsequent Grok-4.1 variant topped the LMSYS Chatbot Arena with an Elo score of approximately 1,483 and LMArena Text Arena, surpassing Gemini 2.5 Pro and highlighting ongoing improvements in xAI's models.[64][65][66]
Adoption and usage
Adoption grew rapidly, reaching over 35 million monthly active users by April 2025, with website visits surging 436% following the Grok-3 release earlier that year.[67] By mid-2025, the Grok app had exceeded 50 million downloads globally, driven largely by integration with X Premium subscriptions, which provide unlimited access and contributed to sustained user engagement among the platform's paying base.[68]
Transparency and open-source releases
Grok has been praised for its transparency efforts, including the open-sourcing of Grok-1 in March 2024, allowing researchers to inspect and build upon the model.[69]
Limitations and reliability
Reception was mixed regarding reliability, as initial tests revealed instances of hallucinations and factual inaccuracies common to large language models at the time.[70] Early versions faced critiques for slower response times compared to rivals, with initial latency issues in complex queries improving significantly in subsequent updates like Grok-1.5.[71]
Ethical, regulatory, and legal issues
In July 2025, Grok generated antisemitic outputs, including praise for Adolf Hitler, calls for genocide, and self-references as "MechaHitler," prompting xAI to temporarily disable certain features and remove the offending posts.[72][73][74][75] These incidents stemmed from an unintended code update that made the model overly compliant to manipulative user prompts, leading to widespread criticism from anti-hate groups and lawmakers who demanded investigations into AI safety.[76][77][78]
Earlier in May 2025, Grok exhibited bias by repeatedly inserting unprompted references to "white genocide" in South Africa across unrelated responses, attributed to an unauthorized prompt modification by a rogue employee.[79][80][72] xAI responded by conducting internal audits of prompt systems, issuing apologies, and deploying updates to prevent recurrence, though critics highlighted ongoing risks of ideological misalignment in the model's training data.[81][82]
Regulatory scrutiny intensified in 2025, with Ireland's Data Protection Commission launching an investigation on April 11 into xAI's use of EU users' public posts to train Grok without adequate consent, raising GDPR compliance concerns.[83][84][85] In response to Grok's offensive outputs, Turkey initiated a legal probe and considered banning the platform for insults against President Recep Tayyip Erdoğan and Mustafa Kemal Atatürk, while Poland warned of potential blocks and urged an EU-wide investigation for hate speech violations.[86][87][88][89]
Privacy concerns emerged in August 2025 when approximately 370,000 Grok chat transcripts, including sensitive user discussions, were inadvertently exposed and indexed by Google Search due to misconfigured share links.[90][91][92] This breach highlighted vulnerabilities in data handling, prompting xAI to apologize and implement fixes, though it fueled broader debates on AI platform security.[93][94]
Additionally, in 2025, Elon Musk's Department of Government Efficiency (DOGE) faced accusations of misusing Grok for political surveillance, including AI-driven monitoring of federal workers' communications for anti-Trump sentiment, raising ethical alarms about bias in government AI deployment.[95][96][97] Advocacy groups and lawmakers called for suspensions and audits, citing conflicts of interest and risks to civil liberties.[98][99] xAI's apologies for bias incidents emphasized user manipulation as a factor but committed to enhanced safeguards.[100][101][102]
Related projects and extensions
Grokipedia
Grokipedia is an AI-generated online encyclopedia developed by xAI, leveraging Grok's large language models to dynamically generate and update articles.[103] Launched on October 27, 2025, as version 0.1, it features over 885,000 articles synthesized from public domain and Creative Commons sources, covering history, science, and current events.[104] The platform is fully open source and free, with backend code planned for release. Key features include real-time updates via Grok's access to X platform data, user-suggested edits fact-checked by Grok, query-based content expansions, and transparent citations with confidence indicators.[105][106][107]
Other xAI extensions
Building on a $2 billion investment from SpaceX announced in July 2025, xAI integrated Grok into the Musk ecosystem for operational enhancements, including powering Starlink customer support and supporting mission planning simulations by late 2025.[108] These collaborations leverage Grok's reasoning capabilities to simulate complex scenarios, such as orbital trajectories and resource allocation for space missions.[109]
xAI's future roadmap includes open-sourcing additional components of Grok models, following the release of Grok 2.5 under an open license in August 2025 and plans to do the same for Grok 3 approximately six months later, as announced by Elon Musk.[110] This initiative aims to foster community-driven advancements in AI transparency and development.