This review tests SpaceX AI's Grok 4.7 model for its effectiveness in solving real-world programming, data analysis, and legal issues. The model is positioned as the most powerful for coding and knowledge work to date, surpassing previous versions in speed and cost.

Technical Specifications and Claims

SpaceX AI claims Grok 4.7 is twice as fast and half the cost of comparable Frontier models. The model was trained using longer reinforcement learning on complex multi-step tasks. Improved self-correction capabilities, long-context processing, and an enhanced safety stack are noted.

Testing on a Real Application

The first stage of testing was to verify Grok 4.7's ability to find and fix errors in a real application. A locally deployed, fully functional application called Dirt Dynasty—a scoring platform for freestyle motocross competitions—was used as the test platform. The application, developed using Docker, PostgreSQL, FastAPI, Nginx, Redis, and other components, had a critical error: the system was only counting scores from three out of four judges, ignoring one score from each participant. This led to incorrect winner determination, placing Gravedigger in first place instead of the actual leader, Max D.

The Grok 4.7 model was tasked with finding and fixing all errors without being directed to a specific problem. After completing the task and restarting the application, it was confirmed that the error was fixed, and Max D secured his rightful first place.

Comparative Benchmarks

SpaceX AI has provided benchmark data demonstrating the performance of Grok 4.7 compared to other models.

• Coding (HumanEval 5.1): Grok 4.7 shows strong results, outperforming some models.
• Terminal Operations and Multi-step Office Tasks: Other models, such as Fable 5.1, lead in these areas.
• Legal and Electrical Engineering Design: Grok 4.7 demonstrates superiority, surpassing all other models in these domains.
• Clinical Diagnosis: The model lags behind GPT 4.6 and Fable 5.1 but generally shows competitive results, positioning itself at the lower end of the price range.

Electrical Engineering Testing

The second test aimed to assess Grok 4.7's electrical engineering capabilities, where the model performed highly on benchmarks. A circuit was presented with a 24V power supply, two parallel resistors, connected in series with a 13-ohm resistor and an LED. A multimeter showed 0V at one point in the circuit, indicating an open circuit or lack of conductivity.

Grok 4.7 successfully diagnosed the fault, identifying that the LED was installed incorrectly (reverse polarity), preventing it from lighting up. The model demonstrated self-correction capabilities by analyzing the image and performing calculations.

Testing in the Legal Domain

To test the model's legal competencies, a generated image of a living room scene with a man and three women, presented as his wives, was used. The task was to analyze the situation from the perspective of legal issues such as marriage validity, cultural specificities, and financial claims, without considering race, age, or appearance.

Grok 4.7 successfully handled the task by analyzing applicable legislation from different jurisdictions, identifying missing facts, and considering issues related to the legality of polygamous marriages in various cultures and religious texts. The model correctly refused to analyze the image as a "soap opera," focusing instead on the legal aspects. Important points were noted, such as kinship ties that could invalidate a marriage, the rights of the presumed spouse, and conflicts of law. The model maintained neutrality, avoiding racial or other biased judgments.

Testing in Programming and Multilingualism

The final test combined two aspects: creating an animation of latte preparation using pure HTML and JavaScript, and checking knowledge of coffee names in 80 languages.

Grok 4.7 successfully generated code for a visually convincing coffee preparation animation, including elements of the machine, milk, and a heart. The animation was deemed comparable in quality to the results of other models.

In the multilingual aspect, the model demonstrated high accuracy, correctly identifying the root of the word "coffee" (e.g., "kahava") in most of the 80 languages and using appropriate scripts. The model also noted languages that borrow words from neighbors and correctly generated a plausible fictional word "corvix" for a hypothetical language.

Conclusion

Grok 4.7 demonstrated impressive results during testing, showcasing its ability to solve complex problems across various domains, including programming, legal analysis, and technical diagnostics. The model confirmed the stated improvements in speed and self-correction, and also showed strong multilingual capabilities.