A model code co-trained with Cursor: Grok 4.5

SpaceXAI launches Grok 4.5, a code and agentic AI model co-trained with Cursor on NVIDIA GB300 GPUs, priced at $2 per million input tokens.

Grok is upgrading to version 4.5. SpaceXAI is positioning this model for code, agentic tasks, and office work, and states it was trained in collaboration with Cursor.

The training involved tens of thousands of NVIDIA GB300 GPUs, with a strong emphasis on data filtering and curation (deduplication, quality scoring, domain selection) and a ramp-up of reinforcement learning focused on multi-step software engineering. The stack is designed for highly asynchronous training, where agentic rollouts run for several hours while learning continues.

Regarding uses, the publisher highlights end-to-end application generation from a single prompt, illustrated by a Three.js solar system simulation produced in one query. Grok 4.5 also becomes the default model for Grok Build and extends to office automation: multi-sheet Excel templates fed by web searches, diagrams built from native shapes in PowerPoint, and writing in Word.

In terms of performance, SpaceXAI serves the model at 80 tokens per second and presents it as economical, with pricing at $2 per million input tokens and $6 per million output tokens.

Access is open today in Grok Build, in Cursor across all offerings, and via the SpaceXAI console, with a temporary free period. The European Union will have to wait: access there is expected by mid-July.