Grok 4.5 (also referred to as Grok 4.5 (high) based on its default reasoning configuration) is a frontier large language model released by SpaceXAI on July 8, 2026. Developed with a primary focus on coding, agentic workflows, and professional knowledge tasks, the model represents a major step in SpaceXAI's engineering-focused AI strategy following its collaborative work with the AI coding environment Cursor. It is designed to deliver fast, highly cost-effective reasoning, positioning itself on the performance-to-cost frontier among contemporary large-scale language models.
Architecture and Training
The model is built upon SpaceXAI's proprietary V9 foundation architecture and is reported to consist of approximately 1.5 trillion parameters in a mixture-of-experts (MoE) configuration, representing a threefold increase in scale compared to its predecessor. Grok 4.5 was trained in SpaceXAI’s Memphis supercluster utilizing developer-agent interactions, codebase telemetry from Cursor, and a broad training mix of high-quality STEM tasks, scientific publications, and professional workplace materials to preserve generalized reasoning capabilities alongside specialized coding functions.
Agentic Capabilities and Performance
A central focus of the model's development was reinforcement learning (RL) in realistic interactive environments. Through this post-training process, Grok 4.5 is optimized to autonomously investigate issues, leverage available software tools, recover from execution errors, and verify final outputs. On agentic evaluations and workplace benchmarks, it demonstrates high efficiency, accomplishing multi-step tasks in fewer steps and using significantly fewer output tokens than comparable models of its tier.
Configuration and Developer Integration
Grok 4.5 features a context window of 500,000 tokens and accepts both text and image modalities as inputs. When querying the API, developers can configure the model's reasoning effort level between low, medium, and high (with high being the default setting). To optimize repeated prompt structures and long-horizon agent interactions, the API supports prompt caching, which discounts cache hits and reduces response latency.