Current image models are terrible at this. (That was tested on DALL-E 2, but DALL-E 3 is no better.)

The image model must get the correct number of sides on at least 95% of tries per prompt. Other details do not have to be correct. Any reasonable prompt that the average mathematically-literate human would easily understand as straightforwardly asking it to draw a pentagon must be responded to correctly. I will exclude prompts that are specifically trying to be confusing to a neural network but a human would get. Anything like "draw a pentagon", "draw a 5-sided shape", "draw a 5-gon", etc. must be successful. Basically I want it to be clear that the AI "understands" what a pentagon looks like, similar to how I can say DALL-E understands what a chair looks like; it can correctly draw a chair in many different contexts and styles, even if it misunderstands related instructions like "draw a cow sitting in the chair".

If the input is fed through an LLM or some other system before going into the image model, this pre-processing will be avoided if I can easily do so, and otherwise it will not. If the image model is not publicly available, I must be confident that its answers are not being cherry-picked.

Pretty much neural network counts, even if it's multimodal and can output stuff other than images. A video model also counts, since video is just a bunch of images. I will ignore any special-purpose image model like one that was trained only to generate simple polygons. It must draw the image itself, not find it online or write code to generate it. File formats that are effectively code, like an SVG don't count either; it has to be "drawing the pixels" itself.