Elon Musk’s SpaceXAI unveiled its next-generation artificial intelligence model, ‘Grok 4.5,’ on the 8th (local time), taking direct aim at the enterprise coding market. The launch came just one day before OpenAI’s scheduled release of its latest model, GPT-5.6, on the 10th, as Musk moved to seize the initiative by claiming the model delivers “Opus-level performance while being faster, more token-efficient, and lower cost.”
The centerpiece of this announcement is not raw benchmark scores but overwhelming ‘token efficiency.’ In the SWE Bench Pro software engineering evaluation test, Grok 4.5 consumed an average of 15,954 tokens per task, while Anthropic’s Claude Opus 4.8 used 67,020 tokens. This roughly 4.2-fold efficiency gap means users can process more than four times the workload for the same cost. Additionally, an inference speed of 80 tokens per second (TPS) delivers response times comparable to lightweight models.
The pricing strategy is equally aggressive. Grok 4.5’s API usage is set at $2 per million input tokens and $6 per million output tokens. This is significantly lower than Anthropic’s Claude Opus 4.8 ($5 input, $25 output) and competitive with OpenAI’s GPT-5.6 Luna ($1 input, $6 output). In the latest evaluation by independent assessor Artificial Analysis, Grok 4.5 ranked fourth globally on the GDPval-AA v2 benchmark, with a per-task execution cost of $0.49 — approximately 90% cheaper than top-tier competitors.
The model was born from a collaboration with Anysphere, the developer of the AI coding tool ‘Cursor,’ which SpaceX acquired for approximately $60.0 billion. It represents the first deliverable just weeks after the acquisition announcement. Grok 4.5 was trained using tens of thousands of Nvidia’s latest GB300 GPUs and underwent large-scale asynchronous reinforcement learning across hundreds of thousands of tasks. A key architectural feature is infrastructure designed to allow agents to sustain uninterrupted learning for hours on complex software engineering tasks.
On the functionality front, Grok 4.5 goes beyond simple code generation to support deep integration with Microsoft Office. It can write complex formulas spanning multiple sheets and add annotations in Excel, or construct charts using basic shapes in PowerPoint — demonstrating an ability to understand and manipulate the structure of real business documents. This is widely interpreted as a strategic foundation for expansion into professional domains such as finance, law, and cybersecurity.
Industry analysts view this launch as a turning point where the axis of AI model competition is shifting from ‘model intelligence’ to ‘real-world workplace dominance.’ Musk has deployed Grok 4.5 for immediate use through the company’s AI coding agent ‘Grok Build,’ Cursor’s developer portal, and the SpaceXAI Console API — effectively constructing a closed ecosystem spanning from model development to distribution within integrated development environments (IDEs). According to Stanford University’s 2026 AI Index Report, as major AI labs’ model performance converges within a narrow band, the competitive focus has already shifted to cost, reliability, and speed.
However, some observers urge caution. Critics note that the ‘Opus-level performance’ Musk referenced is ambiguous regarding which specific version of the Opus model is being compared, and the claim appears based on internal evaluator impressions rather than official benchmark results. AI commentator Jun Song has predicted that Grok 4.5 may perform slightly below the smaller GLM-5.2. Furthermore, while the model’s robust cybersecurity capabilities could aid enterprise security automation, they are also seen as a factor that could attract regulatory scrutiny over potential misuse.
Grok 4.5 is scheduled to launch in the European Union (EU) region starting mid-July. Musk has signaled plans to release a fully retrained new model every month through the end of the year, setting the stage for an even more intense AI model race in the second half of 2026.