Grok 4 - The latest big model from Musk's xAI

What's Grok 4?

Grok 4 is xAI's newest big AI model, Grok 4 delivers a 10x increase in reasoning power compared to its predecessor. The model has superior reasoning capabilities, scoring near perfect on difficult exams such as the SAT and GRE, and outperforming other cutting-edge models in a number of benchmarks. grok 4 supports multimodal functionality, understands subjective concepts, generates code and visualizations, and features significant improvements in voice interaction. grok 4 is available in two versions, Grok 4 Heavy, a single agent version, and Grok 4 Heavy, a multi agent version that supports four agents. Grok 4 is divided into two versions, Grok 4 is a single agent version, and Grok 4 Heavy is a multi agent version, which supports four agents working at the same time, and the context window supports up to 256k tokens.

Main features of Grok 4

• Scientist-level reasoning skills: Trained on xAI's Colossus supercomputer with Ph.D.-level academic problem solving capabilities.
• Deep Knowledge Optimization: To provide more accurate and reliable knowledge by recognizing and correcting misinformation, with the goal of rewriting the human knowledge base.
• multimodal support: Supports text and image input and will be extended to video in the future.
• Advanced Voice Functions: Grok 4 Voice has a natural, real-life voice with end-to-end latency cut in half for a smoother conversation experience.
• Professional Coding ModelGrok 4 Code is optimized for programming, supports multiple languages, efficiently writes, debugs and interprets code, and can be embedded in an IDE to modify code in real time.
• real time web access: Equipped with the DeepSearch tool, which crawls the latest information in real time from web sources such as the X platform.
• Internet Cultural SavvyIt's the most "web-savvy" AI assistant in the world: it understands Internet "stems," slang, and humor with high precision.
• Function Calls and Structured Output: Supports function calls to trigger external tools that return structured data (e.g., JSON) for easy parsing by programs.
• API Support: Available through the xAI API, supporting function calls, JSON mode responses, etc., and compatible with OpenAI and Anthropic.

Grok 4's official website address

• Official website address:: https://x.ai/grok

How to use Grok 4

• Directly via the X platform
  • Subscribe to X Premium+: $16 per month for access to Grok 4's conversational features directly on Platform X (formerly Twitter), with support for real-time web access and image analysis.
  • Certified account privileges: Blue V certified users can be quickly certified through the official website to get the trial qualification on a priority basis.
• Through the SuperGrok app: A standalone application interface that supports voice interaction, file uploads (PDF/Excel, etc.), and in-depth searches for non-technical users.
• Developer API Integration
  • Registration and Key Acquisition
    • Visit the xAI Developer Portal to register for an account.
    • Create API keys, set permissions and rate limits.
  • free quota: New users receive a $150 API credit for the first month (requires participation in the data sharing program).
  • Quick Code Example
    • Python (OpenAI SDK compatible)::

      from openai import OpenAI
      client = OpenAI(
          base_url="https://api.x.ai/v1",
          api_key="YOUR_GROK_API_KEY"
      )
      response = client.chat.completions.create(
          model="grok-4-beta",
          messages=[{"role": "user", "content": "用Python写个快速排序"}]
      )
      print(response.choices[0].message.content)

    • cURL request::

      curl https://api.x.ai/v1/chat/completions \
        -H "Authorization: Bearer YOUR_GROK_API_KEY" \
        -d '{"model":"grok-4-beta","messages":[{"role":"user","content":"分析X平台AI讨论热点"}]}'

Core Benefits of Grok 4

• Interdisciplinary performance at the doctoral level: Exceeds the doctoral level in all major subjects such as mathematics (AIME 25 out of 25), physics, chemistry, and humanities.
• The Ultimate Human Exam Breakthrough: In the HLE benchmark test covering 2,500 PhD-level puzzles, Grok 4 Heavy (multi-agent mode) became the first model in the world to pass the halfway mark with an answer rate of 50.7%.
• AGI Test Leadership: Record score of 15.81 TP3T on the ArcAGI v2 test, which is close to the standard for general-purpose AI, twice as much as the second-place finisher (Claude Opus).
• Grok 4 Heavy Multi-Agent Collaboration: It supports parallel reasoning of 4 intelligences, and improves the efficiency of complex problem solving exponentially through cross-validation and scheme optimization.
• Real-time dynamic optimization: MLB Championship probability prediction was completed in just 4.5 minutes in the demo, integrating information retrieval, data modeling and probability computation.
• arithmetic crushing: Based on Colossus supercomputer (200,000 GPU cluster), training computation is 100 times higher than Grok 3, and response speed is increased by more than 50%.
• First token generation delay: 10 seconds in 32K tokens context, 15% faster than Grok 3.
• special coding model: Grok 4 Code supports one-click embedding into IDEs (e.g. Cursor), and code generation accuracy and efficiency exceeds that of GPT-4 Code Interpreter.
• API Automation: It supports function calls, JSON structured output, and can automatically trigger external APIs, which is suitable for high-precision scenarios such as finance, law, and healthcare.
• Cost Advantage: $3 per million tokens input and $15 per million output, only 1/3 the cost of Claude 3 Opus.

Model testing for Grok 4

• official test::
  • Humanity's Last Exam: Contains 2,500 cross-disciplinary, expert-level questions.Grok 4 Heavy scores 44.41 TP3T with the tool, and can be increased to 50.71 TP3T with further optimization.
  • AIME25 (math competition): The Grok 4 Heavy got a perfect score of 100%, crushing all other models.
  • GPQA (Graduate Proficiency Question and Answer): Grok 4 Heavy scored 88.91 TP3T, ahead of the Gemini 2.5 Pro (86.4%) and Claude 4 Opus (79.6%).
  • HMMT25 (High School Math Competition): Grok 4 Heavy scored 96.71 TP3T, well ahead of Gemini 2.5 Pro (82.51 TP3T).
  • USAMO25 (United States Mathematical Olympiad): Grok 4 Heavy scored 61.91 TP3T, significantly ahead of Gemini DeepThink (49.41 TP3T) and Gemini 2.5 Pro (34.51 TP3T).
  • ARC-AGI (Abstract Reasoning): Grok 4 scored 15.91 TP3T, nearly doubling the previous commercial SOTA.
  • Vending-Bench (simulation): Grok 4 netted $4694, well ahead of Claude Opus 4 ($2077) and Human Player ($844).
• Third-party evaluation(Artificial Analysis test, a platform for evaluating the performance of large models):
  • Artificial Intelligence Index (AI): Grok 4 picked up 73 points, outperforming OpenAI o3 (70), Google Gemini 2.5 Pro (70), Anthropic Claude 4 Opus (64) and DeepSeek R1 0528 (68 points).
  • Coded and mathematical indices: Grok 4 were both ranked first.
  • GPQA Diamond score: Record high of 88%, surpassing the Gemini 2.5 Pro's 84%.
  • Humanity's Last Exam Score: Record high of 24%, surpassing the Gemini 2.5 Pro's 21%.
  • tempo: Grok 4 at 75 tokens/sec, not as good as o3 (188 tokens/sec) and Gemini 2.5 Pro (142 tokens/sec), but better than Claude 4 Opus Thinking (66 tokens/sec).

Product Pricing for Grok 4

• paid subscription program::
  • SuperGrokThe annual fee is $300 and the monthly fee is $30.
  • SuperGrok HeavyThe annual fee is $3,000 and the monthly fee is $300.
• API Call Pricing::
  • importation: $3 / million tokens.
  • exports: $15 / million tokens.

Who Grok 4 is for

• Top Developers: Full-stack engineers, algorithm experts, and open source project maintainers who need to handle multi-million code bases or build complex systems.
• AI/research workers: college professors, lab researchers, and data scientists for academic breakthroughs, experimental simulations, or interdisciplinary analysis.
• technology entrepreneur: Startup CTOs, independent hackers, need 48 hours to validate a product from 0 to 1 or automate operations.
• Financial Quantitative Team: Hedge funds, high-frequency trading organizations that rely on real-time data and PhD-level reasoning to develop strategies.
• National/enterprise-level institutions: Heavy R&D in aerospace, energy, pharmaceuticals, etc., requiring private deployments to solve ultra-complex engineering problems.