Grok 4.5 (1.4T) goes live in Cursor; early users report strong results

Cursor rolled out Grok 4.5 on July 8 and said the model is now available across its desktop app, web app, iOS app, CLI and SDK. The company described it as a model built jointly with SpaceXAI and positioned it as broader than a coding-only release, saying it was designed for “difficult, long-running tasks” across software engineering and other computer-based work. (cursor.com) That matters because the story is not just that Grok 4.5 “went live in Cursor.” It is that Cursor is treating it as a first-party model in its own product stack, bundling usage into subscription plans and giving customers double usage for the first week of launch. (cursor.com) Cursor also said it added new safeguards tied to the model’s cybersecurity capabilities. ### So what exactly launched inside Cursor? Cursor said Grok 4.5 is available “today in Cursor” and can be selected across the company’s main surfaces, including desktop, web, iOS, CLI and SDK. (cursor.com) The launch blog says the model was trained jointly with SpaceXAI and is the company’s “most intelligent model” as well as the first one it built for more than software engineering. Cursor’s documentation page describes the release in similar terms, saying Grok 4.5 is aimed at long-running coding and knowledge-work tasks and listing pricing and availability details alongside the model page. (cursor.com) ### What does Cursor say the model is built for? Cursor said Grok 4.5 is intended for tasks that require creative tool use and recovery from mistakes in realistic environments. In its launch post, the company said the training process used reinforcement learning on difficult problems spanning both software engineering and broader knowledge work. (cursor.com) The company also said the model was trained on trillions of tokens of Cursor data, which it said captured user interactions with codebases and software tools. Cursor said that broader training mix included STEM tasks, research papers and other knowledge-work material rather than only coding-specialist data. (cursor.com) ### What numbers are actually public? Cursor said Grok 4.5 is a mixture-of-experts model, but the company’s public launch materials reviewed here do not state a parameter count. The 1.4 trillion-parameter figure circulating in social posts was not independently confirmed in the official Cursor sources reviewed for this article. (cursor.com) Cursor did publish token pricing. The company said the base model costs $2 per million input tokens and $6 per million output tokens, while a fast variant costs $4 per million input tokens and $18 per million output tokens. (cursor.com) Cursor also said individual and team plans include usage from its first-party model pool. ### Are there any benchmark caveats behind the early enthusiasm? Cursor’s own CursorBench page shows Grok 4.5 High at a 66.7% score with an average cost per task of $1.51, alongside Medium at 65.4% and Low at 63.5%. On the same page, Cursor says Grok 4.5 had “an advantage” because an earlier snapshot of the Cursor codebase was unintentionally included in training, and it says the exact score impact is unclear. (cursor.com) That disclosure is important because it puts a qualification on one of the main pieces of quantitative evidence attached to the release. (cursor.com) Cursor said the affected data has been removed for future models. ### What are users likely to watch next inside the product? Cursor staff said in a July 9 community forum reply that Grok 4.5 currently has a 256,000-token context window inside Cursor. The same reply said users may be able to access a different context window through the SpaceXAI API, but “in Cursor” the limit is 256k. (cursor.com) Cursor’s forum and docs are also where the company has been clarifying billing and rollout details. In a July 8 reply, a Cursor staff member said Grok 4.5 usage draws from the renamed “First Party Models” pool after a brief launch-period attribution issue. (cursor.com) (forum.cursor.com 1) (forum.cursor.com 2)