SpaceXAI releases Grok 4.5, co-trained with Cursor

Background

Grok is the family of AI models and chatbot products built by xAI. The release described here is Grok 4.5, a model for coding, agentic tasks and knowledge work.

What happened

On 16 July 2026 the company announced Grok 4.5, which it called its most capable model to date. It is aimed at coding, agentic tasks and knowledge work^[1]. The model was trained on tens of thousands of NVIDIA GB300 GPUs and was co-trained with Cursor^[1].

Key details

According to the company, Grok 4.5 scored 64.7% on SWE-Bench Pro and 83.3% on Terminal Bench 2.1^[1]. These are xAI's own figures. API pricing was $2 per million input tokens and $6 per million output tokens^[1].

The model became the default in Grok Build, and it was offered in Cursor on all plans^[1]. Grok Build itself was released under the Apache License on the same day^[2].

Two days earlier, on 14 July 2026, Grok Build had been reported to upload full user repositories, including secrets, to cloud storage. According to Wikipedia, that behaviour was changed on 14 July as well^[2].

Impact

The release put Grok 4.5 into two coding environments used by developers: as the default model in Grok Build and as an option for all Cursor plans^[1].

For readers new to these terms, a benchmark is a fixed set of test tasks used to compare models. A percentage such as 64.7% is the share of those tasks the model handled successfully under that benchmark's rules. API prices are charged per million tokens, so a request with 10,000 input tokens at $2 per million costs about $0.02 for its input. The Apache License is a common open-source software licence that allows others to use and modify the code under its terms. For a reader, the main points are the date, the model's stated focus, the training partner, and the two ways it reached users: as the default in Grok Build and as an option in Cursor^[1].