SpaceXAI Launches Grok 4.5: A Powerful Model for Coding and Knowledge Work SpaceXAI has officially launched Grok 4.5, its latest and most advanced model designed to excel in coding, agentic tasks, and knowledge work. The model is positioned as the company’s strongest offering to date, developed in tandem with Cursor and built to handle complex engineering challenges with enhanced efficiency and intelligence. Grok 4.5 was trained on a diverse dataset encompassing coding, science, engineering, and mathematics, with a focus on real-world applications. The model’s performance was benchmarked against leading competitors in tasks such as DeepSWE, SWE Marathon, and Terminal Bench 2.1, where it demonstrated strong capabilities. For instance, in the DeepSWE 1.0 benchmark, Grok 4.5 achieved a pass@1 score of 66.1%, outperforming models like Opus 4.8 max (40.12%) and GPT 5.5 xhigh (62.0%). These results highlight its proficiency in software engineering and technical problem-solving. The training process for Grok 4.5 involved scaling reinforcement learning techniques, with a focus on per-token intelligence. The model was trained across tens of thousands of NVIDIA GB300 GPUs, utilizing advanced data filtering and curation methods. This included deduplication, quality scoring, and domain-specific selection to ensure high-coverage and high-signal data. The training stack was optimized for asynchronous operations, enabling agentic tasks to run efficiently over extended periods. Grok 4.5’s capabilities extend to creative and practical applications. For example, the model can generate fully functional applications from a single prompt, such as creating a simulation of the universe and solar system using Three.js. The simulation includes adjustable time, realistic motion, and a well-designed HUD aligned with modern design principles.#cursor #spacexa_i #grok_4_5 #deep_swe #terminal_bench_2_1
