I think GPT-3 is used as a benchmark for performance, so saying a model is on par with GPT-3 should give you an idea of what you can get out of it. IIRC most o

I think GPT-3 is used as a benchmark for performance, so saying a model is on par with GPT-3 should give you an idea of what you can get out of it. IIRC most open source models to date - including the semi-open LLaMAs - have GPT-3-like performance. Nothing gets close to GPT-3.5 and beyond.