Not a game dev but found out qwen3-coder:30b is a good balance for speed and quality for coding.

Almost don't use perplexity or chatgpt. When qwen3 fails with wrong answer, all others do the same.

But found out sometimes it is better than those online "services". Sure it is way better than gpt-oss:20b and that was a surprise.

Most of the times I just want to get sintaxe or code snippets.

It works also very well with long code and many programs from the same application.

So it runs on Ollama and I use OpenWebUI. It is very fast, way faster then gpt-oss:20b and it runs on old hardware AMD Ryzen 3600x with 10gb 3080 and 32gb of ram.

Never used ChatGPT 5 but the older failed too many times for me. Perplexity is ok, but I end up using search engines myself because the answers were not correct and many times very dangerous.

I'm a game and game engine dev and I use Cursor as my daily driver. I rarely exceed my $20/mo plan. It's been very helpful as a pair programmer and I've developed an optimized workflow for using it with my engine architecture.

To be fair, I think all the models and tools are essentially functionally equivalent, like word processors or washing machines. People report the same kind of good results with command line tools like Claude Code.

As regards the game industry and tech, it's often not innovative. Yes, of course there are breakthrough graphic developments and such, but game implementation is all ugly spaghetti under the hood. It's an industry that's slow to adopt new tech and tooling and AI is in that bucket.

I also think the art side of the game business strongly objects to generative AI on principle and, on art quality, in practice. This anti-AI stance can bleed over into the dev team and prevent them from fully embracing the coding work that it's very good at.