Grok 4.5, the first joint Cursor and SpaceXAI model

Cursor and SpaceXAI released Grok 4.5 on July 8, 2026 — the first model the two companies have built together, and Cursor's first model aimed at more than software engineering. The release lands weeks after SpaceX agreed to acquire Cursor in a deal reported at $60 billion, making Grok 4.5 the first visible product of that combination.

Definition: Grok 4.5 is a mixture-of-experts model jointly trained by Cursor and SpaceXAI, built for software engineering, data science, finance and legal work rather than coding alone.

Example: Instead of only writing and debugging code, Grok 4.5 can review a contract, analyze a financial statement, or research a regulation, using the same tool-calling loop it uses inside a codebase.

Key takeaway: This is Cursor's first model trained on a deliberately broad task mix rather than a coding-only dataset, and the first released since SpaceX's acquisition of Cursor.

Business impact: Teams already running Cursor for engineering get access to a model that claims to extend into adjacent knowledge work — worth testing against a real legal, finance or research task, not just a coding benchmark.

What is Grok 4.5, and who built it?

Grok 4.5 is a mixture-of-experts model that Cursor says it trained jointly with SpaceXAI, described in Cursor's own announcement as the company's "most intelligent model and the first we've built for more than software engineering." Cursor frames Grok 4.5 as capable of handling "difficult, long-running tasks that require creatively using tools to solve problems," extending beyond code into data science, finance and legal work. SpaceXAI is the AI unit behind Elon Musk's Grok model line and its Colossus compute infrastructure, which Cursor first began leveraging for model training through an April 2026 partnership — before SpaceX moved to acquire Cursor outright two months later. More on this: Cursor Releases Cursor Router: A Request-Level Classifier Delivering Frontier Coding Quality at 30–50% Lower Cost.

Why did Cursor build this with SpaceXAI instead of on its own?

Cursor's own coding-specialist models — Composer, Composer 1.5, Composer 2, and Composer 2.5 — were bottlenecked by compute, which is the reason Cursor gave for partnering with SpaceX on training in the first place, to tap SpaceXAI's Colossus infrastructure. According to The Tech Portal, SpaceXAI trained Grok 4.5 across tens of thousands of Nvidia GB300 GPUs — a training scale Cursor could not reach with its own compute. Elon Musk described the result to The Tech Portal as "roughly comparable to Opus 4.7, but much faster," adding that "hardcore engineers at Tesla & SpaceX find Grok 4.5 genuinely useful, which is what actually matters" — a framing that puts real internal usage ahead of benchmark scores.

What data did Grok 4.5 actually train on?

Grok 4.5's training mix is the clearest technical break from Cursor's earlier Composer models. Cursor's announcement says training included trillions of tokens of Cursor's own usage data, capturing real interactions between developers and codebases as well as between coding agents and their environments — plus a deliberately broad mix of STEM tasks, research papers and other knowledge work, rather than the coding-only mix used to train Composer 2.5. Yahoo Finance reports that Cursor, which has roughly 7 million monthly users across more than 50,000 engineering teams, supplied that real-world developer data as supplemental training on top of xAI's existing Grok foundation model. This is the same underlying pattern behind how AI agents learn to route tasks: a model that investigates a problem, chooses a tool, and verifies its own output rather than answering in one shot.

How did SpaceXAI test Grok 4.5 on hard problems?

Cursor's announcement describes training Grok 4.5 with reinforcement learning on problems deliberately built to be difficult enough that even frontier models fail at them, spanning both software engineering and broader knowledge work — teaching the model to investigate a problem, use tools, recover from its own mistakes, and verify results before finishing. To build enough of these environments, Cursor says it used a distributed agent system: engineers specify a problem and how a correct solution is verified, then large groups of agents construct, test and refine each environment at scale — work Cursor says would otherwise have taken teams of hundreds of engineers months to complete. Cursor frames this as using its previous model generation to accelerate training of the next one.

How does Grok 4.5 score against Opus 4.8, GPT-5.5 and Fable 5?

Cursor's own launch chart puts Grok 4.5 in the middle of the pack on its four published benchmarks, not ahead of every rival model:

Claude Fable 5 leads three of the four benchmarks, and Claude Opus 4.8 leads on SWE-Bench Multilingual — a model whose own SWE-bench Pro and Terminal-Bench scores Anthropic has published for its Sonnet line too, for comparison. The "high," "max" and "xhigh" labels next to some scores mark different reasoning-effort settings per model, so DeepSWE 1.0 and SWE-Bench Pro results aren't run at identical compute budgets across the table — a detail worth checking before treating any single row as a clean apples-to-apples comparison.

What does Grok 4.5 add for legal and finance teams specifically?

Beyond coding, The Tech Portal reports that Grok 4.5 is built to review contracts, summarize case files, compare legal documents and analyze regulations on the legal side, and to support financial statement analysis, investment research, due diligence, market intelligence and risk assessment on the finance side. Cursor's own announcement is less specific, describing the model's reach only as "software engineering, data science, finance, legal work, or anything else you do on a computer" — so the concrete task list for non-coding use cases currently comes from SpaceXAI's public statements rather than Cursor's own post.

Where can I use Grok 4.5, and what does it cost?

Grok 4.5 is live today across Cursor's desktop, web, iOS, CLI and SDK products. Cursor's individual and team subscription plans include what Cursor calls "significant usage" of the model as part of its first-party model pool, with usage doubled for the first week after launch. Outside a subscription, the base model is priced at $2 per million input tokens and $6 per million output tokens, and a faster variant costs $4 per million input tokens and $18 per million output tokens. Cursor also says it added new safeguards specifically reflecting Grok 4.5's cybersecurity capabilities, without detailing what those safeguards are.

What should teams watch next with Grok 4.5?

Cursor's own post flags one open question directly: Grok 4.5 shows an advantage on CursorBench because an earlier snapshot of Cursor's codebase was accidentally included in its training data. Cursor says the exact impact on that score is unclear, has removed the contaminated data for future models, and is working on a larger CursorBench revision — which is why CursorBench doesn't appear in the comparison table above at all. Composer 2.5 also isn't going away: Cursor says it and Grok 4.5 are different model weight classes, and it will keep offering Composer-sized models going forward. Grok 4.5's API availability for outside developers hasn't shipped yet either — Cursor is only taking signups for that release. The open question for teams already on Cursor is less "does Grok 4.5 win the leaderboard" and more whether its broader task range holds up on a real legal, finance or data-science workflow rather than a coding benchmark. Related reading: Grok 4.6 Targets Long-Running Agents and Visual Work.