xAI has officially launched Grok 4.5, describing it as the company's most intelligent model to date, purpose-built for coding, agentic tasks, and knowledge work. The new model was developed in collaboration with Cursor and represents a significant step forward in real-world engineering capabilities.
Training at Massive Scale with Curated Data
Grok 4.5 was trained on datasets covering coding, science, engineering, and mathematics. The training process leveraged tens of thousands of NVIDIA GB300 GPUs, supported by stability techniques specifically designed for large-scale operations.
Beyond sheer computational volume, xAI placed heavy emphasis on data quality. The company implemented deduplication, quality scoring, and domain-focused selection to ensure the training data remained both high-coverage and high-signal. This meticulous curation process was aimed at maximizing the value of every token fed into the model.
Reinforcement learning was scaled with a particular focus on per-token intelligence. The RL training encompasses hundreds of thousands of tasks, primarily centered on multi-step software engineering and other technical work. Both automated and model-based grading systems were employed to evaluate performance.
The training stack was engineered for highly asynchronous operations, allowing agentic rollouts to run for extended periods, sometimes many hours, while learning continued simultaneously across the full GPU cluster. This architecture ultimately produces more intelligent and efficient reasoning on complex engineering and agentic tasks, according to the company.
Strong Coding and App-Building Capabilities
xAI reports that Grok 4.5 demonstrates exceptional proficiency in coding, handling challenging tasks in Rust and C/C++ as well as end-to-end application development from a single prompt. The model can create well-designed, fully functional applications even when given minimal specifications.
As an illustrative example, the company highlighted the model's ability to generate a detailed solar system and universe simulation using Three.js, complete with adjustable time controls, realistic orbital motion, stars, and a modern heads-up display, all from a single prompt.
Speed, Efficiency, and Competitive Pricing
Grok 4.5 is served at 80 tokens per second, matching the speeds typically associated with fast models. Combined with approximately twice the token efficiency of comparable leading models, it delivers results more quickly and at significantly lower cost.
The model is priced at $2 per million input tokens and $6 per million output tokens. xAI notes that Grok 4.5 solves tasks in under half the number of steps required by competing models, resulting in what the company describes as the highest intelligence per unit of time and cost.