Cursor's Grok 4.5 was built by AI agents, not humans. That's the real story.

On Monday, Cursor announced the launch of Grok 4.5, a Mixture-of-Experts model it trained jointly with SpaceXAI. But the headline, a smarter model with higher benchmark scores, misses the real story. The most significant development isn't the model itself; it's how it was built.

The self-reinforcing loop

Grok 4.5 is the first model whose training environments were constructed at scale by autonomous AI agents. According to Cursor, engineers define a problem and a verification method, then large groups of agents build, test, and refine each training environment. Some of these problems, the company says, would have required months of work from teams of hundreds of engineers to create manually.

"It's one of the ways we used the previous model to accelerate progress of the next model," the company wrote in its announcement.

This is a self-reinforcing loop: each generation of models helps build the environments that teach the next generation. The result is an accelerating flywheel of capability, where the rate of improvement compounds.

More than a coding specialist

Unlike Composer 2.5, Cursor's previous model that was optimized narrowly for coding, Grok 4.5 retains a broader training data mix. The company deliberately included high-quality STEM tasks, research publications, and other forms of intellectual work. The result is a model that claims mastery across software engineering, data science, finance, law, and any computer-based profession.

Grok 4.5's training started with trillions of tokens from Cursor's interaction data, capturing how developers and agents work across codebases and tools. It then underwent reinforcement learning on hard problems in realistic environments, learning to analyze, use tools, recover from mistakes, and verify results.

The company notes that as models improve, existing tasks stop teaching anything new, problems that once required deep reasoning become trivial. The new training pipeline is designed to keep generating sufficiently hard problems.

Benchmarks and a data contamination caveat

Cursor published benchmark scores for Grok 4.5 on SWE-Bench Pro, Terminal-Bench, and a multilingual SWE-Bench. But the company also disclosed that Grok 4.5 benefits from an advantage on CursorBench: an earlier snapshot of Cursor's codebase was accidentally included in training. "The exact impact is not clear," Cursor writes. The data has been removed from future models, and Cursor is working on a larger update to CursorBench.

This transparency is rare in an industry that often brushes data contamination under the rug. It also means CursorBench scores should be taken with salt.

Pricing and availability

Grok 4.5 is available today on Cursor's desktop, web, iOS, CLI, and SDK. Individual and team plans include significant usage in the proprietary model pool, with usage doubled during the first week. Pricing for the base model is $2 per million input tokens and $6 per million output tokens. A faster variant costs $4 per million input and $18 per million output.

Composer 2.5 remains available as a separate weight class, and Cursor says it will release more models in that class in the future.

The bigger picture

Cursor's approach, using AI agents to build training environments for the next generation of models, may be more consequential than any single model release. If this self-reinforcing loop scales, it could dramatically compress the timeline for achieving general capability across a wide range of intellectual work.

The company's playbook is clear: instead of competing on model size alone, compete on the efficiency of building better models using the models you already have. And that might be Grok 4.5's most important lesson.