SpaceXAI introduced Grok 4.5 - a new model for programming, agent tasks and intelligence that is positioned as the company’s "strongest system". Available via API, Grok Build and Cursor on all plans (EU launch expected mid-July). The main focus is on cost and speed: $2/$6 per million of inputs/outputs compared to $5-10/$25-50 for Claude Opus 4.8, Fable 5 and GPT-5.6 Sol.

Test results are mixed: Grok 4.5 leads SWE Marathon (29%) and Harvey’s Legal Agent Benchmark, but lags Fable and GPT-5.5 behind DeepSWE 1.1 and SWE Bench Pro. Key advantage - efficiency: 15,954 output tokens per task on average vs. 67,020 for Claude Opus 4.8 (4.2 times less). At a lower price, this lowers the total cost for scenarios with multiple iterations.

The model was trained in conjunction with Cursor using developer sessions (debugging tracings, real code edits). SpaceX entered into a deal to purchase Cursor for $60 billion in class A shares (closing - Q3 2026). Tens of thousands of Nvidia GB300 GPUs are used to train Grok 4.5.