Grok 4.5 Shocked Me. Here’s Why.

It seems like every week, a new model hits the headlines. But SpaceXAI’s latest release took me completely by surprise.

For context, SpaceXAI’s last model, Grok 4.3, was not even considered to be a true frontier model, scoring only 38 on the Artificial Analysis Intelligence Index, while its competitors, GPT 5.5 and Claude 4.7 Opus, scored 55 and 54, respectively. At that point, I honestly had no expectations for Grok’s later models, but I was caught off-guard when I saw Grok 4.5 show up in OpenRouter.

The first thing I saw was the price: just $2/M input tokens and $6/M output tokens. But this pricing comes with a catch: when you have more than 200k tokens in context, the price doubles to $4/M input tokens and $12/M output tokens.

While the price to run an AI model does go up the more context you give to it, it doesn’t increase that dramatically. Probably, this is because the model costs that much from the start, but Grok wants users to start using it at a cheaper price.

The second thing that I saw was how good the model was. It scored 54 on the Artificial Analysis Intelligence Index, which is on par with the last generation of AI models. It also scores a 76 on the Artificial Analysis Coding Agent Index using Grok Build, tying it with GPT 5.5 using Codex. It’s also incredibly efficient token-wise, so it’s third in terms of output tokens per task.

Benchmarks are good, but nothing tells you how good a model is better than testing it myself. I gave it the entire codebase for the desktop app of AcceleratedLogic AI, which was at that point just a UI, and made Grok 4.5 add the ability to connect AI models to it, and it did, in under 100k tokens. I continued to use it to edit the code, and found it to excel at programming in a similar way to Opus-class models.

The question on most people’s minds is probably “how did Grok get so good?” From its first model release in 2023, to controversies in 2025, Grok seemed like it never would truly be at the frontier. But one acquisition changed all of that. On June 16, 2026, SpaceX acquired Cursor, a popular agentic coding app for $60 Billion, after months of being partners. That sounds like a waste, especially considering that Cursor operates at a huge loss, but SpaceXAI saw that Cursor had something invaluable: user data. If you have enough data, and the proper training technique, along with enough compute, you could create a powerful AI model, and the team at SpaceXAI knows this. To get all of the coding data from Cursor, they updated the terms of service on June 28, allowing for real world developer data to flow to SpaceXAI’s supercomputer.

In the end, Grok 4.5 hasn’t completely dominated the leaderboards, and definitely isn’t the best AI model out there, but it does prove something: that not just OpenAI, Anthropic, and Google can make frontier AI models.