SpaceXAI has launched Grok 4.5, its most capable model yet. It is the company’s first release since going public and buying the AI coding startup Cursor. This is the joint model the two firms had raced to ship. Elon Musk aimed it squarely at coding and agentic work, not casual chat.

“It is an Opus-class model, but faster, more token-efficient and lower cost,” Musk wrote on X. The line name-checks Anthropic’s top Opus family. A chart with the announcement claims Grok 4.5 beats Opus 4.8 on several benchmarks, Axios first reported.

Built for coders, and for Wall Street

Grok 4.5 was trained alongside Cursor, which SpaceX agreed to buy in a deal valuing the startup at $60bn. The model is built to “handle difficult, long-running tasks,” according to the company’s blog post, including software engineering. Unlike Cursor’s earlier models, it also targets legal and financial work, and adds cybersecurity features, Bloomberg reported.

The finance push is deliberate. Musk said this year that his AI unit, known as xAI before it merged with SpaceX, had fallen behind on coding. The company has since rebuilt the team and chased Wall Street clients for its Grok chatbot. Grok 4.5 is the clearest sign yet of that pivot towards paying business customers.

The price play

Cost is the pitch. SpaceXAI priced Grok 4.5 at $2 per million input tokens and $6 per million output tokens. Anthropic’s Opus 4.8 runs at $5 and $25, while OpenAI’s GPT-5.6 Luna sits at $1 and $6. That undercuts the priciest rivals as companies watch their token spend more closely.

The company concedes limits. It says Grok 4.5 beats some OpenAI and Anthropic models on speed, price and performance. It does not beat their largest and latest. Musk expects to close that gap soon. The model is live now in Grok Build, in Cursor on all plans, and via the SpaceXAI console. A wider public release is due Thursday. It is not yet available in the EU.

Why it matters

The release lands on a crowded day. OpenAI is rolling out GPT-5.6 widely on Thursday, alongside a new set of voice models. That comes after the Trump administration asked it to stagger the launch. Government scrutiny hangs over all of it, with regulators watching new models for cybersecurity risk. Cursor said it has taken steps to “detect and block bad actors” while preserving legitimate security research.

There is a twist in how Grok 4.5 was built. SpaceXAI trained it on the same compute it leases to rivals Anthropic and Google. As its own models grow hungrier, it will have to choose. It can feed them, or rent that capacity out for cash. For now, Musk is betting a cheaper, coder-friendly Grok can win business, even while it trails the best models on raw power.