SpaceXAI, the AI division under SpaceX, announced and released its latest model, Grok 4.5, on July 8, 2026. Developed in collaboration with the AI coding tool Cursor, the model is designed for software development, multi-step agentic tasks, and knowledge work. SpaceXAI identifies this as the company's most high-performance model to date.

According to API documentation, Grok 4.5 features a 500K token context window and supports both text and image inputs. Pricing is set at $2 per 1M input tokens, with cached inputs priced at $0.5 and output tokens at $6. SpaceXAI reports an inference speed of 80 tokens per second, claiming the model completes tasks with fewer tokens compared to major competitors.

In performance evaluations, the model achieved a 29.0% resolution rate on the SWE Marathon benchmark for long-duration software development, surpassing Opus 4.8 at 26.0% and Fable at 24.0%. In the DeepSWE 1.0 benchmark, the model reached 62.0%, trailing behind Fable and GPT-5.5. These benchmarks were self-reported by SpaceXAI, and the company notes that environmental variations mean these may not represent identical testing conditions.

The training process utilized tens of thousands of NVIDIA GB300 GPUs to process datasets across coding, science, engineering, and mathematics. Capability in multi-step software development was refined through reinforcement learning on hundreds of thousands of tasks. This is the first flagship model for SpaceXAI, which changed its name from xAI on July 6. The model is scheduled for release in the EU in mid-July, with a separate pricing structure for requests exceeding 200K tokens.

SpaceXAI, the AI division under SpaceX, announced and released its latest model, Grok 4.5, on July 8, 2026. Developed in collaboration with the AI coding tool Cursor, the model is designed for software development, multi-step agentic tasks, and knowledge work. SpaceXAI identifies this as the company's most high-performance model to date.

According to API documentation, Grok 4.5 features a 500K token context window and supports both text and image inputs. Pricing is set at $2 per 1M input tokens, with cached inputs priced at $0.5 and output tokens at $6. SpaceXAI reports an inference speed of 80 tokens per second, claiming the model completes tasks with fewer tokens compared to major competitors.

In performance evaluations, the model achieved a 29.0% resolution rate on the SWE Marathon benchmark for long-duration software development, surpassing Opus 4.8 at 26.0% and Fable at 24.0%. In the DeepSWE 1.0 benchmark, the model reached 62.0%, trailing behind Fable and GPT-5.5. These benchmarks were self-reported by SpaceXAI, and the company notes that environmental variations mean these may not represent identical testing conditions.

The training process utilized tens of thousands of NVIDIA GB300 GPUs to process datasets across coding, science, engineering, and mathematics. Capability in multi-step software development was refined through reinforcement learning on hundreds of thousands of tasks. This is the first flagship model for SpaceXAI, which changed its name from xAI on July 6. The model is scheduled for release in the EU in mid-July, with a separate pricing structure for requests exceeding 200K tokens.