On July 16, 2026, SpaceXAI (the former xAI, now under SpaceX) launched Grok 4.5, which it calls its smartest model, built to excel at coding, agentic tasks, and knowledge work. Unusually, the model was trained alongside Cursor, the AI coding editor, and ships as the default model in SpaceXAI’s Grok Build product as well as in Cursor on all plans and via the SpaceXAI API console.
Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs, with reinforcement learning scaled over hundreds of thousands of tasks centered on multi-step software engineering, using a highly asynchronous stack that lets agentic rollouts run for many hours while learning continues. On SpaceXAI’s published benchmark charts it scores 62.0 percent on DeepSWE 1.0 (behind Claude Fable 5 max at 66.1 and GPT 5.5 xhigh at 64.31, ahead of Claude Opus 4.8 max at 55.75), leads SWE Marathon at 29.0 percent, and posts 83.3 percent on Terminal Bench 2.1.
The pitch is efficiency rather than outright leadership: the model is served at fast-model speeds of 80 tokens per second and resolves SWE Bench Pro tasks with about 15,954 output tokens on average, roughly 4.2x fewer than Opus 4.8 max. Pricing is $2 per million input tokens and $6 per million output tokens, undercutting frontier-tier rivals and landing a week after OpenAI’s GPT-5.6 family in a period of intense price-performance competition. SpaceXAI is also positioning Grok 4.5 for office work, with plugins that drive Excel, PowerPoint, and Word natively.