Space XAI has released Grok 4.5, its latest AI model designed for coding, agentic tasks, and knowledge work. Trained on extensive datasets covering coding, science, engineering, and mathematics, Grok 4.5 demonstrates strong reasoning capabilities. The model achieved the top score on Harvey's Legal Agent Benchmark and performed competitively across various software engineering benchmarks, including Deep SWE 1.0 and Terminal Bench 2.1.

Grok 4.5 was trained on tens of thousands of NVIDIA GB300 GPUs, with significant investment in data filtering and curation for high-signal quality. The training process focused on reinforcement learning for multi-step software engineering tasks, utilizing an asynchronous training stack that supports long-running agentic rollouts. This approach aims for more intelligent and efficient reasoning, enabling the model to generate complex applications and code from single prompts.

Beyond its coding prowess, Grok 4.5 excels in office applications like Excel, PowerPoint, and Word, capable of creating complex models, diagrams, and clear prose. Delivered at 80 TPS with enhanced token efficiency, Grok 4.5 aims to provide faster, lower-cost intelligent results. It is available via API, in Grok Build, and within Cursor, with pricing set at $2/million input tokens and $6/million output tokens.