SpaceXAI officially released Grok 4.5 on July 8, positioning it as their most powerful AI model to date, designed for code writing, agentic tasks, and advanced knowledge work; the highlight of this release is its joint training with the well-known AI code editor Cursor. Benchmark results include a 64.7% success rate on SWE Bench Pro.
Grok 4.5 Trained with Cursor: 64.7% on SWE Bench Pro
According to SpaceXAI, Grok 4.5's performance on various benchmarks is as follows:
SWE Bench Pro: 64.7% success rate
Terminal Bench 2.1: 83.3% (GPT 5.5 scores 83.4%, a 0.1 percentage point difference)
DeepSWE 1.0: 62.0%
Harvey's Legal Agent Benchmark: ranked first
SpaceXAI states that Grok 4.5 covers hundreds of thousands of multi-step software engineering tasks through large-scale reinforcement learning (RL), supporting long-duration and asynchronous agent operations. Additionally, Grok 4.5 can build complete applications from a single prompt, such as generating a 3D universe and solar system simulator with HUD interface using Three.js.
Powered by Tens of Thousands of NVIDIA GB300 GPUs, Inference Speed Reaches 80 TPS
Based on official data, Grok 4.5 is built on the computing power of tens of thousands of NVIDIA GB300 GPUs, achieving an inference speed of 80 tokens per second (TPS). In terms of token efficiency, it is more than twice as efficient as comparable models for the same tasks; in SWE Bench Pro testing, Grok 4.5 outputs approximately 15,954 tokens on average, reducing consumption by 4.2 times compared to Opus 4.8.
The official pricing is $2 per million input tokens and $6 per million output tokens. SpaceXAI states this offers the highest "intelligence per unit of time and cost."
Microsoft Office Integration: Writing Long Articles in Word, Modeling in Excel
According to SpaceXAI, Grok 4.5 is now fully integrated into the Microsoft Office ecosystem via official plugins. Users can write long articles in Word, handle web research and complex formulas in Excel, and automatically generate native-shape charts in PowerPoint.
Grok 4.5 is now the default model for "Grok Build." Developers in most regions worldwide can obtain API keys through console.x.ai and start integration. Due to local regulatory restrictions, access in the EU has not yet been opened; SpaceXAI plans to officially launch EU access in mid-July.
Frequently Asked Questions
What are Grok 4.5's main benchmark scores?
According to SpaceXAI, Grok 4.5 achieves 64.7% on SWE Bench Pro, 83.3% on Terminal Bench 2.1 (GPT 5.5 scores 83.4%), and 62.0% on DeepSWE 1.0; it also ranks first on Harvey's Legal Agent benchmark.
What is the API pricing for Grok 4.5?
As per SpaceXAI, the API costs $2 per million input tokens and $6 per million output tokens; in SWE Bench Pro testing, it averages about 15,954 tokens output, with token efficiency more than twice that of similar models.
When will Grok 4.5 be available in the EU?
Due to local regulatory policies, access in the EU is currently restricted. SpaceXAI plans to officially open EU access in mid-July 2023.
