Comment by ddtaylor

I noticed that LLMs like to write code and anytime an "AI feature" is needed they will heavily default to using `gpt-4o` as kind of the "hello world" of models. It was a good model when it came out and a lot of people started building on it, which caused the training data to be saturated by it.

My AGENTS.md has:

    You MUST use a modern but cost effective LLM such as `qwen3-8b` when you need structured output or tool support.

The reality is that almost all LLMs have quirks and each provider tries their best to smooth them over, but often you might start seeing stuff specific to OpenAI or the `gpt-4o` model in the code. IMO the last thing you want to be doing in 2026 is paying higher costs to use an outdated model being kept on life support that needs special tweaks that won't be relevant once it gets the axe.