A typical (in fact, the modal) line from the database, which goes back to April, is "oyakadon 1/2c rice, 2 thighs, 1/2 breast, 2 eggs, bunch broccolini, std oyakadon broth", and the model works out for itself that these are average-sized thighs and breast (boneless and skinless, because it's oyakadon). "std oyakadon broth" is a reference to a "house recipe", of which there's a side database, each costed by GPT5. There's a low/med/high adjuster on every entry; sometimes the thighs are bigger or smaller, for instance.
I've:
* Checked these against all three major frontier models and gotten the same macros breakdown within relatively tight bands
* Manually researched a couple of the most common recipes
* Run this system over many months against my most likely TDEE and reconciled against scale weight loss (which has been profound).
My confidence in frontier model's ability to generate "good enough, probably more than good enough" from this level of meal description is quite high.
Key simplifying assumption (for the model): I'm cooking most of my meals. I cook. I virtually never order out or eat packaged food of any sort. This would be a harder problem and the answers less reliable if I was giving it, like, "1/2 JP Graziano muffaletta, extra hot giardiniera".
You can actually get just as good results for packaged food if you just allow the agent to use a web search tool & keep a growing local database of "known food items". I did the same for a while, but it was quite expensive (this was a while back with less powerful models and required a bit of orchestration to make it reliable)
Presumably it would Just Work if I wrote "1/2 bag cool ranch doritos", which is one of the virtues of this input/processing system.