Grok has been thoroughly unimpressive for the past few years, but after xAI purchased Cursor, version 4.6 is really back in the game. It now competes with the top models in coding quality, knowledge depth, and most other metrics, and it's the least expensive of the current multimodal frontier models ($2/M input tokens $6/M output tokens). It also performs fast.
We're talking about benchmarks that beat Fable in some cases.
Here's a 1-shot result for the prompt 'Please create an html 3d car racing game':
The only other model which created a driving game with similar vertical motion (hills) was Kimi K3, and that required many more prompt iterations, was much more expensive to complete and took far longer to create. This is a high end, extremely smart model.
There have been so many amazing model releases in the past few weeks. This one is definitely worth paying attention to. It's not open source, but should at very least be explored as a potential top tier tool in your hosted API collection.