Both are fine, though frontier models are more persistent. They mostly take unwarranted shortcuts, trying to find the easiest way to reach the literal goal set by you, failing to see the broader picture. But they definitely do not confabulate, as in they don't just make up nonexistent things, they would rather walk extra mile to verify when they're not certain, which is the reason the software they make is trash but it works in the narrow sense. They can be rigid in their writing style due to mode collapse, but I haven't seen this leading to them inventing nonexistent functions or anything like that in a long while (and it's a job for formal tools to catch that, in any case).

Modern LLMs are very far from being next token predictors trained only on text. Text and other modalities is the base that makes them understand your language and have concepts similar to yours, but they were also subjected to RL that makes them good at reaching the goals. Two tricks in a trenchcoat instead of just one.