SpaceXAI releases Grok 4.5, the first model trained jointly with Cursor

SpaceXAI, the company formed by merging Elon Musk's xAI with SpaceX, released the new Grok 4.5 model on July 8. It's the first launch since SpaceX announced in June that it was acquiring Cursor for $60 billion in stock, and Grok 4.5 is the first fruit of that merger, trained in part on data from Cursor coding sessions.

Elon Musk described the new model briefly on X, comparing it directly to Anthropic's strongest model before the release of Opus 4.8. Rather than chasing the competition on raw benchmark scores, SpaceXAI is betting on a combination of speed, lower price, and token efficiency.

It's an Opus-class model, but faster, more token-efficient, and cheaper - Elon Musk, SpaceXAI

First fruit of the Cursor merger

Cursor, the popular AI-assisted coding tool with roughly 7 million monthly users and more than 50,000 engineering teams, was acquired by SpaceX in June for $60 billion in stock, just after the record-breaking stock market debut of Musk's rocket company. The deal was expected to close in the third quarter of the year, and the two companies had spent months jointly training a model on xAI's Colossus computing infrastructure.

Grok 4.5 is the result of that collaboration. The model was trained on trillions of tokens of real-world data from Cursor coding sessions, supplemented with material from STEM, scientific research, and knowledge work. As a result, Cursor can market the model not just as a coding tool, but also for tasks in data science, finance, and law.

Pricing and performance

At launch, Grok 4.5 costs $2 per million input tokens and $6 per million output tokens in the SpaceXAI API, with a lower rate of $0.5 for tokens read from cache. Inside Cursor, a fast variant is available priced at $4 per million input tokens and $18 per million output tokens, where the premium buys priority handling rather than a different model. The context window is 500,000 tokens, and the model responds at a speed of about 80 tokens per second.

In terms of architecture, Grok 4.5 is a mixture-of-experts model estimated at around 1.5 trillion parameters, a fifty percent jump in scale over Grok 4.4 in under a month. Training ran on tens of thousands of Nvidia GB300 chips. The model accepts text and images, returns text, and supports a reasoning_effort parameter, structured responses, and native parallel tool calling.

Law and finance as a new testing ground

In benchmark tests, Grok 4.5 took first place on the Harvey Legal Agent Benchmark, a test measuring models' ability to handle real legal work. SpaceXAI highlights use cases in contract review, case file summarization, comparing legal documents, and regulatory analysis. In finance, the model is meant to support financial statement analysis, investment research, due diligence, and risk assessment.

In coding benchmarks, the results are more mixed. On the SWE Marathon test, Grok 4.5 achieved a 29 percent solution success rate, ahead of Opus 4.8's 26 percent, but on several other tests, such as DeepSWE or SWE Bench Pro, Fable and GPT-5.5 models remain in the lead. SpaceXAI is therefore betting more on cost and speed than on topping every individual benchmark.

Europe has to wait

The model is not yet available to users in the European Union, and the company says access will launch in mid-July without giving a specific reason for the delay. For Polish developers using Cursor, this means the new, cheaper model will reach their instances several weeks behind the US market, even though Cursor has long been one of the more popular AI coding tools in Poland.

For software companies and law firms weighing whether to adopt AI tools, the key comparison will be Grok 4.5's pricing against what Anthropic and OpenAI offer. At $2 per million input tokens, the model comes in noticeably cheaper than both competitors' flagship models, which could put pricing pressure on the entire API market in the coming months.

The consolidation of SpaceX, xAI, and Cursor under a single SpaceXAI brand also points to a new strategy for Musk: rather than competing on the language model alone, the company is building a full stack, from computing infrastructure through the model to the developer tool where that model actually works day to day.

Sources: SpaceXAI, Cursor Launch Grok 4.5 AI Model for Finance, Legal Applications (bloomberg.com), Introducing Grok 4.5 (x.ai), SpaceXAI releases Grok 4.5, which Elon describes as an 'Opus-class model' (techcrunch.com), SpaceX to acquire the AI coding startup Cursor for 60 billion (cnbc.com), SpaceXAI Ships Grok 4.5 With Cursor, EU Access Delayed (aiweekly.co)