>>96426677Claude has terrible context compared to other models if you're talking about anything more complex than NIAH and finding information like that.
There was a context test Nvidia put out called RULER that basically tested and saw that none of the LLMs are good past 32k and significantly degrade despite the context listed. The latest research test in this area shows that too, context is terrible for more complex things and a LLM has terrible performance. For Claude, that's around the 4k mark and 8k-16k while usable is where it goes off the rails. 32k unusable per the test.
