It’s called “genAI Capabilities Canvas”
It lists 18 capabilities of LLMs and helps think through which ones are needed for a genAI application.
https://x.com/genAIArch/status/1703416032839487488?s=20
As I see it GPT-4 is clearly the most capable today…
If you need “tool use” you need to use GPT models because those are the only ones specifically trained for LLM function calling.
If you want “styling” to be done by fine-tuning then again GPT3.5 and soon GPT4 is the strongest to start with.
However, if you’re not using that capability there may be other, cheaper options.