Grok 4.5 Lands on Cursor — A Jointly-Trained Coding Model That Undercuts Opus on Price

On July 8 and 9, 2026, SpaceXAI rolled out Grok 4.5, the company's most intelligent model so far. It is the first frontier model SpaceXAI has trained specifically for software engineering and agent tasks. Training ran on tens of thousands of NVIDIA GB300 GPUs, in partnership with the AI coding company Cursor.

The joint training combined Cursor's billion-scale user interactions with SpaceXAI's compute footprint. Grok 4.5 was built for tasks that span hundreds of skills and tools, from refactoring code across multiple repositories, to building complex Excel models, drafting Word documents, and designing PowerPoint slides. Office workloads are now treated like software engineering.

The price turned heads. The API is set at 2 dollars per million input tokens and 6 dollars per million output tokens, roughly half of comparable premium models. On the SWE-Bench Pro coding benchmark, Grok 4.5 finishes each task with about 15,900 output tokens, while a leading rival needs roughly 67,000. The model streams at 80 tokens per second.

The launch comes weeks after SpaceX agreed to acquire Cursor in a deal worth around 60 billion dollars. Grok 4.5 is now the default inside the Grok Build console, with usage limits reset for early testers, and is expected to reach the European Union by mid-July. Analysts read the move as Musk's clearest signal yet that he plans to win on cost and speed rather than raw model size.