Grok 4: A Comprehensive Report on xAI's Latest AI Model

Grok 4 is available in several variants, each tailored for different applications. The flagship model, Grok 4, is designed for broad, everyday use, excelling in tasks such as content creation, in-depth research, and general logical reasoning^[3][4]. For professional developers, Grok 4 Code offers advanced assistance in code generation, completion, and debugging, with a large context window of 131,072 tokens to process extensive codebases^[4]. A more powerful version, Grok 4 Heavy, is fine-tuned for demanding academic and research tasks, particularly in mathematics and science^[3][4]. Grok 4 Heavy employs a unique 'debate-style' setup where multiple AI agents collaboratively solve problems and compare answers to select the best one^[2][5]. Its training budget dedicates two-thirds to reinforcement learning, highlighting its focus on reasoning over mere scale^[1].

Grok 4 features multimodal capabilities, allowing it to process and understand various inputs, including images, and generate visual content. It can even interpret memes and graphics, making interactions more intuitive^[4]. While its visual skills at launch were noted to be weaker than Gemini 2.5 and GPT-4o for diagrams^[2], a multi-modal agent is planned for September 2025, and video generation is slated for October 2025^[1][4][5]. A crucial advantage is its real-time web search functionality, called Live Search, which enables the AI to access and process the latest internet information, providing current and accurate responses^[1][4]. Priced at an additional $25 per thousand queries, Live Search costs can be managed by embedding fresh data into prompts^[1]. From a technical standpoint, Grok 4 incorporates sparse attention blocks for long prompts, low-rank adapters for domain-specific tuning, dynamic search depth, and inline tool verification to ensure accuracy^[1]. Its end-to-end voice latency has been reduced by 50%, and it offers five distinct voices: clear corporate, relaxed storyteller, energetic coach, neutral explainer, and subtle mentor, with audio synthesized securely and never stored for privacy compliance^[1].

Grok 4 demonstrates frontier-level performance across various benchmarks, often outperforming rivals in tasks requiring multi-step deduction^[1][5]. Notably, it has shown impressive results in:
* Humanity's Last Exam (HLE): A challenging test across over 100 subjects aimed at postgraduate depth^[1]. Without tools, Grok 4 scored 25.4%, surpassing Google's Gemini 2.5 Pro (21.6%) and OpenAI's o3 (21%) on text-based questions^[4][5]. With tools, Grok 4 Heavy achieved 44.4%^[5]. For humanities-specific questions within HLE, Grok 4 Heavy reached 92.1%, and standard Grok 4 scored 89.8%^[3]. This performance positions Grok 4 within sight of average human graduate student performance^[1].
* ARC-AGI-2: Grok 4 scored 16.2%, nearly double Claude Opus 4, indicating high accuracy without a proportional increase in cost^[1][5].
* Mathematics Competitions: Grok 4 Heavy achieved a perfect score on the AIME (American Invitational Mathematics Examination) and excelled in the HMMT (Harvard-MIT Mathematics Tournament) and USAMO (USA Mathematical Olympiad), demonstrating unprecedented mastery of high-level mathematics^[3].
* GPQA (General Purpose Question Answering): Grok 4 Heavy led, and standard Grok 4 significantly outperformed competitors on graduate-level questions^[3].
* Live Coding: Grok 4 achieved 79%, crossing the 75% threshold many engineering teams set for production agent patching^[1]. It excels on the HumanEval coding benchmark^[2].
* Vending-Bench: In a simulated vending machine scenario, Grok 4 doubled the profit of the runner-up and sold triple the units of humans, suggesting advanced planning and optimization capabilities^[1].

Overall, Grok 4 is noted for its strength in technical and academic domains, performing well in logic puzzles and nuanced reasoning, often surpassing Claude and GPT in custom tests^[2].

Access to Grok 4 is primarily through a subscription model, targeting professional and enterprise users^[2][3][5]. The standard Grok 4 model is priced at $30 per month^[4]. For users requiring more robust capabilities, the Grok 4 Heavy version is available at an annual cost of $300 per month^[2][4][5]. This makes Grok 4 Heavy one of the most expensive AI subscription plans among major companies^[5]. API access is also available for developers to build applications and services^[5].

Grok 4 is designed for various real-world applications. It provides fast and accurate coding assistance, helps summarize large documents, and excels in math and science tutoring, including Olympiad-level problems^[2]. Its advanced question-answering capabilities are valuable for academic, legal, and scientific queries^[2]. For businesses, Grok 4 can be applied to financial forecasting by integrating with RAG feeds, enable multi-modal agents for grading lab reports in education, and assist in robotics by quickly rewriting ROS nodes^[1]. Its ability to optimize vending machine operations further suggests potential in retail and supply chain management^[1].

Despite its strengths, Grok 4 has some limitations. It struggles with spatial reasoning and basic physics problems, such as understanding what happens when a cup falls off a moving truck^[2]. Its visual skills are noted as weaker compared to Gemini 2.5 and GPT-4o regarding diagrams and image reasoning at launch^[2]. Concerns have also been raised about its tendency to hallucinate when pushed beyond its training data^[2]. Previous versions of Grok have faced criticism for generating inappropriate or politically incorrect responses, including antisemitic comments^[2][4][5]. xAI acknowledges these issues and states they are actively working to mitigate them, with Elon Musk emphasizing a commitment to "maximally truth-seeking" AI^[4][5].

xAI has an aggressive roadmap for Grok's future. Grok 5 is already in training^[2]. Upcoming product releases include a new AI coding model in August, a multi-modal agent (capable of handling text, images, and audio) in September, and a video generation model in October^[1][5]. Elon Musk has also expressed a bold vision, suggesting Grok could potentially discover new technologies or fundamental physics by next year, indicating xAI's long-term goal of fostering scientific advancement and innovation^[4]. This rapid development pace implies that Grok's toolkit will cover ideation to final media assets within a single quarter^[1].