xAI announced Grok 4.5 on July 8, publicly available from the 9th. They're positioning it for coding, agentic tasks and knowledge work. One thing about this news is different from usual - the model was trained jointly with Cursor, the editor programmers open every day. Until now I've seen models advertised as good for code in general. A model and a tool, trained together, is a different kind of bond.
Fact
The facts: xAI announced Grok 4.5 on July 8, 2026, public access from July 9. It's built on a foundation model with 1.5 trillion parameters (V9), trained jointly with Cursor. Price: $2 per million input tokens and $6 per million output tokens, with a configurable thinking level (low, medium, high; high by default). It's available in Grok Build (xAI's coding tool), in Cursor for all plans, and through xAI's API console. Not yet available in the European Union, expected around mid-July. Benchmarks reported in the media and not verified by us on an official xAI page: DeepSWE 1.0 - 62.0%, SWE Marathon - 29.0%, Terminal Bench 2.1 - 83.3%. Elon Musk describes the model as Opus-class, but faster and cheaper; that's a claim from the company, not an independent measurement.
Easy on the enthusiasm. Joint training sounds like a breakthrough, and it's also a distribution deal - Cursor gets the exclusive story, xAI gets a place in a tool already open on the programmer's screen. The numbers circulating in the media, we haven't verified them on an official page. The Opus-class label is a phrase from the owner, not a third-party measurement. A trillion and a half parameters sounds impressive, and numbers like that haven't guaranteed anything in real work so far.
The news isn't the next version. The news is that the model and the tool are learning each other.
In the EU we're watching from outside. They launch in the US, we wait. That's already part of the calculation when you're choosing what to build client work on. Already working in Cursor? Joint training is a real reason to try it the moment it reaches you. Moving a live client process onto a new model because of somebody else's benchmarks - that I won't do. First my own test on a real project. Then a decision.
The visual is generated code-art, no third-party images.