SpaceXAI launches Grok 4.5 model on Grok Build and APIs

SpaceXAI released Grok 4.5, a flagship model for coding and agentic workflows, now live in Grok Build, Cursor, and the SpaceXAI API. The model features a 500,000-token context window, function calling, and configurable reasoning, with API pricing at $2 per million input tokens. SpaceXAI claims the model achieves 62% on DeepSWE 1.0 and 83.3% on Terminal Bench 2.1, trained on tens of thousands of NVIDIA GB300 GPUs.

SpaceXAI has released Grok 4.5 https://www.testingcatalog.com/spacexai-gearing-up-for-upcoming-grok-4-5-release/ , its new flagship model built for coding, agentic workflows, and knowledge work. The model is live now in Grok Build, Cursor on all plans, and the SpaceXAI API console, with EU access not yet active and expected in mid-July. The launch puts Grok 4.5 directly into developer workflows rather than positioning it only as a chatbot upgrade. SpaceXAI says the model was trained alongside Cursor and is now the default model in Grok Build, where it can operate as a coding agent through the CLI, terminal UI, scripts, bots, and Agent Client Protocol integrations. It also runs inside Office add-ins for Word, PowerPoint, and Excel, where the company says it can create spreadsheet models, slide diagrams, and document drafts. Grok 4.5 supports text and image input, text output, a 500,000-token context window, function calling, structured outputs, web search, X search, code execution, and configurable reasoning. Developers can set reasoning effort to low, medium, or high, with high used by default and reasoning not disableable. API pricing is listed at $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens, with higher-context pricing applying above 200,000 tokens. SpaceXAI is pitching the model around real engineering tasks. In its published benchmarks, Grok 4.5 scored: - 62.0% on DeepSWE 1.0 - 53% on DeepSWE 1.1 - 83.3% on Terminal Bench 2.1 - 64.7% on SWE Bench Pro The company also claims the model serves at 80 tokens per second and uses 15,954 output tokens on average for SWE Bench Pro tasks, which it frames as about 4.2 times fewer tokens than Opus 4.8 max in the same comparison. The model was trained across tens of thousands of NVIDIA GB300 GPUs, with data filtering, deduplication, quality scoring, domain-focused selection, and reinforcement learning over hundreds of thousands of tasks. SpaceXAI https://www.testingcatalog.com/tag/grok/ says the RL process targeted multi-step software engineering and technical work, with automated and model-based grading plus long-running agentic rollouts.