Elon Musk's SpaceXAI officially launches its most powerful model, "Grok 4.5"! Partnering with Cursor to aggressively target AI coding agents, with perfect integration into Office software.

AI Agent War Gets a Major Disruptor! SpaceXAI today (8th) announced the launch of its most powerful model, "Grok 4.5." Co-trained with Cursor, this next-generation model is designed for coding and agentic tasks. It not only tops multiple real-world engineering benchmarks but also seamlessly integrates with Microsoft Office. With an astonishing inference speed of 80 TPS and exceptional cost-effectiveness, Grok 4.5 vows to redefine enterprise productivity.

(Background: Musk says Grok 4.5 launches tomorrow: Compared to Opus, faster and cheaper, going head-to-head with GPT-5.6) (Background supplement: What's the political stance of your AI? ChatGPT is the most left-leaning, Grok is the only right-leaning, and the most neutral model is...)

Table of Contents

Toggle

• Co-trained with Cursor, dominates real-world engineering tests
• Powered by tens of thousands of GB300 GPUs, breakthrough in inference speed and cost-effectiveness
• Strong integration with Microsoft Office; EU expected to open in mid-July

In the white-hot battlefield of generative AI moving toward physical office applications and autonomous agents, SpaceXAI has once again dropped a bombshell.

On July 8, 2026, Taiwan time, SpaceXAI officially released its most powerful AI model to date — Grok 4.5. Tailored specifically for coding, agentic tasks, and high-level knowledge work, this model heralds the evolution of AI from mere text assistance to an all-capable virtual colleague that can autonomously solve complex engineering problems and deeply participate in business processes.

Co-trained with Cursor, dominates real-world engineering tests

One of the biggest highlights of Grok 4.5 is its co-training with the renowned AI code editor Cursor. The company states that Grok 4.5 demonstrates unparalleled dominance in real-world software engineering, science, and mathematics, not only surpassing comparable leading models but also delivering impressive scores in multiple authoritative benchmarks:

• SWE Bench Pro: Achieved an ultra-high resolution rate of 64.7%.
• Terminal Bench 2.1: Scored 83.3%, closely following GPT 5.5's 83.4%.
• DeepSWE 1.0: Scored a high 62.0%.

Additionally, Grok 4.5 boasts the ability to "build a complete application from a single prompt." Developers only need one sentence, and it can use Three.js to generate a 3D universe and solar system simulator with a modern HUD interface, capable of precisely adjusting orbits and time, significantly lowering the threshold for product prototyping.

Powered by tens of thousands of GB300 GPUs, breakthrough in inference speed and cost-effectiveness

Behind this performance beast lies a massive computing infrastructure based on tens of thousands of NVIDIA's most advanced GB300 GPUs. SpaceXAI has invested significant resources in data filtering, deduplication, and quality scoring, and through large-scale reinforcement learning (RL), it covers hundreds of thousands of multi-step software engineering tasks, enabling perfect support for long-running, asynchronous agent operations.

In terms of performance, Grok 4.5's inference speed has surged to an astonishing 80 TPS (Tokens Per Second). Even more noteworthy is its token efficiency: under the same tasks, it is more than twice as efficient as comparable models, averaging only about 15,954 tokens output in SWE Bench Pro tests — a 4.2x reduction in consumption compared to the previous Opus 4.8. Paired with a friendly pricing of $2 per million input tokens and $6 per million output tokens, the company proudly claims this offers the highest "intelligence per unit time and cost."

Strong integration with Microsoft Office; EU expected to open in mid-July

Beyond hardcore programming, Grok 4.5 also extends its reach to general office administrative and document processing, and it achieved first place in the Harvey’s Legal Agent benchmark. It has become the default model for "Grok Build" and is seamlessly integrated into Microsoft's Office ecosystem through official plugins. Users can now write long articles directly in Word, process models involving web research and complex formulas in Excel, and even automatically generate native shape charts in PowerPoint.

Currently, developers in most regions can access API keys via console.x.ai and start integrating Grok 4.5. However, SpaceXAI also reminds that due to local regulatory policies, users in the EU region are not yet enabled, with access expected to officially open in mid-July.