Member-only story
Grok 4.5 is the fourth-smartest model in the world, and that is not the interesting number. On Artificial Analysis’s independent Intelligence Index it scores 54 and lands fourth, behind Fable 5, GPT-5.5, and Claude Opus 4.8. It wins exactly two of the four benchmarks SpaceXAI chose to publish about it. By the usual way we rank these things, it is a very good model that is not the best at anything.
Here is the number that matters. On SWE-Bench Pro, xAI reports Grok 4.5 resolves a task using an average of 15,954 output tokens. Opus 4.8, on the same benchmark at its maximum setting, burns 67,020. That is 4.2 times fewer tokens for a task both models can solve. Now put the price beside it: Grok 4.5 charges $6 per million output tokens, Opus 4.8 charges $25. Multiply it out and the same solved task costs about $0.096 of output on Grok and $1.68 on Opus. Seventeen times cheaper. Not because Grok is smarter. Because it shuts up sooner.
That is the whole story of the model that SpaceXAI and Cursor shipped on 8 July 2026, and it is a story about economics, not leaderboards.
Key takeaways #
Grok 4.5 launched 8 July 2026, the first model co-trained by SpaceXAI (the merged SpaceX and xAI) and Cursor after SpaceX’s $60 billion acquisition of Cursor’s parent…