SpaceXAI and Cursor just released Grok 4.5, a powerful new mixture-of-experts model that Elon Musk describes as an "Opus-class" alternative at roughly one-third the cost of Anthropic's top-tier offering. Trained jointly with Cursor on trillions of tokens of developer-agent interaction data, Grok 4.5 is designed to handle long-running, tool-intensive tasks across software engineering, data science, finance, and legal work. With base pricing at $2 per million input tokens and $6 per million output tokens, it is the most aggressive pricing we have seen from a frontier model lab in recent memory.

What Makes Grok 4.5 Different

Grok 4.5 is not just another incremental model refresh. It represents a strategic shift for SpaceXAI and Cursor in two important ways.

Broader training beyond code. Unlike Cursor's previous Composer 2.5 model which was a coding specialist, Grok 4.5 was trained on a deliberately broader dataset that includes high-quality STEM tasks, research papers, and other knowledge work. This gives it proficiency across domains well beyond software engineering.

Reinforcement learning on realistic environments. The model was trained using reinforcement learning on difficult, real-world problems that span both software engineering and general knowledge work. These environments teach the model to investigate problems, use tools, recover from mistakes, and verify results. SpaceXAI and Cursor designed a distributed agent system to construct these training environments at scale, with engineers specifying problems and verification criteria while large groups of agents built, tested, and refined each environment.

Pricing that reshapes the market. The base Grok 4.5 model costs $2 per million input tokens and $6 per million output tokens. A fast variant is available at $4 per million input tokens and $18 per million output tokens. These figures undercut every comparable frontier model on the market today.

Grok 4.5 vs the Competition: Pricing Breakdown

Here is how Grok 4.5's pricing compares to other leading models:

Model	Input Price (per MTok)	Output Price (per MTok)
Grok 4.5 (base)	$2	$6
Grok 4.5 (fast)	$4	$18
Anthropic Opus 4.7/4.8	$5	$25
Anthropic Fable 5	$10	$50
GPT-5.6 Sol	$5	$30
GPT-5.6 Luna	$1	$6
GPT-5.4	$2.50	$15

Grok 4.5's base tier is competitive with OpenAI's budget Luna offering on output tokens, while significantly faster and more capable. Against Opus 4.7, Grok 4.5 delivers roughly comparable capability at roughly 60 percent lower cost with faster inference.

What Developers Can Expect

Early user reports on Hacker News paint a nuanced picture. One developer who tested Grok 4.5 against GPT-5.5 and Claude for building an iOS bike computer app found that Grok performed best, successfully producing a native SwiftUI+Metal implementation while Claude fell back to an HTML/CSS workaround.

Another HN commentator noted the economics: "It seems to be extremely economical -- 4x better reasoning efficiency compared to Opus while being priced at $2/$6."

However, Grok 4.5 is not yet available to European users, and the model falls just short of best-in-class on formal benchmarks according to SpaceXAI's own published metrics. The key value proposition is the combination of competitive capability, faster speed, and dramatically lower cost.

How Grok 4.5 Was Built

The model is a mixture-of-experts architecture trained jointly by SpaceXAI and Cursor. The training dataset included trillions of tokens of Cursor telemetry data, capturing a wide range of developer-agent interactions across real codebases. This gave the model an unusually rich understanding of how developers actually work with AI tools in practice.

Cursor used a distributed agent system to build training environments at scale. Some of these training environments would have taken teams of hundreds of engineers months to build using traditional methods. This recursive approach, where the previous model helped accelerate progress on the next one, is a recurring pattern in frontier AI development.

Where to Use Grok 4.5

Grok 4.5 is available today in Cursor across desktop, web, iOS, CLI, and SDK. Individual and team plans include significant usage of the model as part of Cursor's first-party model pool, with usage doubled for the first week.

SpaceXAI plans to make Grok 4.5 available to the broader public today, following positive feedback from beta test customers. The model is positioned as a workhorse capable of handling coding, app building, research, writing, and routine knowledge work.

Frequently Asked Questions

How does Grok 4.5 compare to GPT-5.6 Sol? OpenAI's GPT-5.6 Sol is priced at $5/$30 per MTok and is described as OpenAI's strongest model yet. Grok 4.5 is priced at $2/$6 per MTok and aims to deliver comparable capability at lower cost. Both launched within days of each other in a busy week for AI releases.

Is Grok 4.5 available in Europe? Not yet. Early reports indicate European users may face restrictions at launch, though wider availability is expected to follow.

Can Grok 4.5 replace Claude Opus? According to Elon Musk, Grok 4.5 is "roughly comparable to Opus 4.7, but much faster." For teams prioritizing cost efficiency and speed, Grok 4.5 is a compelling alternative. For specialized use cases where Opus 4.8 or Fable 5 excels, Anthropic's models may still hold an edge.

What tasks is Grok 4.5 best at? The model excels at long-running, tool-intensive tasks that require creative problem solving. This includes software engineering, data science, finance, legal work, and complex research tasks.

How is Grok 4.5 different from Composer 2.5? Composer 2.5 was a coding specialist. Grok 4.5 is a broader, more capable model that handles knowledge work across many domains. Both model weight classes will continue to be offered.

Key Takeaways

• Grok 4.5 is a joint release from SpaceXAI and Cursor, priced at $2/$6 per MTok
• The model is an MoE architecture trained on trillions of tokens of developer-agent interaction data
• It is roughly comparable to Opus 4.7 in capability, but faster and significantly cheaper
• Available now in Cursor across desktop, web, iOS, CLI, and SDK with double usage for the first week
• Not yet available to European users
• Part of the busiest week in AI model launches, alongside OpenAI's GPT-5.6 Sol and GPT-Live

Conclusion

Grok 4.5 represents a serious entrant in the frontier AI model market. Its aggressive pricing, broad domain capability, and tight integration with Cursor make it an attractive option for developers and teams looking to reduce AI costs without sacrificing capability. As SpaceXAI's first major launch since going public, Grok 4.5 signals that the company intends to compete on both performance and price, and that combination could reshape how developers think about model selection in the second half of 2026.

---

Sources: Cursor Blog - Introducing Grok 4.5, TechCrunch - SpaceXAI releases Grok 4.5, SiliconANGLE - SpaceXAI's Grok 4.5 undercuts on price, Hacker News Discussion