Disclaimer: I'm always running on max and don't really have token limits so I am in a position not to care about cost per token. But I am not surprised by the improved benchmark results at all, 4.6 was really not nearly as strong of a model as people seem to remember it being.

I don't find that LLMs are any more likely than humans to remember to update all of the places it wrote redundant functions. Generally far less likely, actually. So forgive me for treating this claim with a massive grain of salt.

This is especially obvious when viewing the reasoning trace for models like Claude, which often spends a lot of time speculating about the user's "hints" and trying to parse out the intent of the user in asking the question. Essentially, the model I use for LLMs these days is to treat them as very good "test takers" which have limited open book access to a large swathe of the internet. They are trying to ace the test by any means necessary and love to take shortcuts to get there that don't require actual "reasoning" (which burns tokens and increases the context window, decreasing accuracy overall). For example, when asked to read a full paper, focusing on the implications for some particular problem, Claude agents will try to cheat by skimming until they get to a section that feels relevant, then searching directly for some words they read in that section. They will do this even if told explicitly that they must read the whole paper. I assume this is because the vast majority of the time, for the kinds of questions that they are trained on, this sort of behavior maximizes their reward function (though I'm sure I'm getting lots of details wrong about the way frontier models are trained, I find it very unlikely that the kinds of prompts that these agents get very closely resemble data found in the wild on the internet pre-LLMs).