If you have not spent the last 1.5 months with Astra or Opus 5.5 in daily dev you may have a skewed view of what their capabilities are based on social media reading. I'll break down why there is such a big wave happening in games as a result of these specific models, which are substantially better than what had come before. Coding a game requires an understanding of the language you are coding in, the structure of the code you want to write, the logic by which you lay down the code, the architecture it sits in, the components it connects to / affects, and if you are writing in an engine, the engine and its editor and plugins. Earlier this year some of these components- code structure and language- were good, but the LLMs lacked good logic, components, and editor understanding. Plus many of the systems lacked good MCP support; UE didn't add a lot of theirs until July, for example, and even then it required a lot of engine restarts. In just the last few months most of this had been fixed. Astra was the first model, launched in the second half of August, where I felt like it had the understanding of Unreal at the level of a mid-career programmer, but it blew away tokens faster than Scarface. Then we got Opus 5.5 in the second half of September and it had the capabilities of Astra 6.1 but at 1/5 the token cost, meaning I could go several days before hitting my weekly usage cap. Not only did the model properly understand how Unreal, Godot, Threejs and other systems worked, but the editors actually played nice with them. More importantly, the logic and structure of their code was nearing senior dev, and if you put it at the highest settings, could actually develop strong architectures, though it could still jump to wrong assumptions if it misunderstood you. Anyone complaining about bugs is probably misunderstanding an actual bug for incorrect logic. I find it very, very rarely makes standard bugs (like forgetting a character or putting the wrong string in, something that my best devs still do all the time). The logic issues will get better as time goes on and the models self correct. They will always be limited by an understanding of what you want to build, but they are superior at the speed by which they can code and test their own coded. Concurrent with this the recent models became excellent at Blender usage. This means they can adjust models and animation themselves, without you needing to do a lot of the legwork. This does not mean they are great at them yet; I find its animation quality to be at the level of a junior animator. It still lacks understanding of key body parts and smooth movement. But the path to getting there is clear and I suspect this will be solved within six months. They are also still bad at lighting and visual/gameplay fidelity feel. I suspect lighting it will get better at, I think gameplay feel may take a long time. These components combined are what has brought us to our current point.