I see Anthropic have now rolled out their 1M token context window in Claude Code. No increase in price, even though the context window is held on expensive servers not the PC. Looking forward to trying it next week, I haven't had time yet to see what difference it makes.
If the "how much it understands and remembers and knows how to really take into account" scaled linearly with context size, such increase would be huge. Clearly 200k is kind of limiting. It limits how much of "generic knowledge" you are supposed to load in every time (e.g. CLAUDE.md) not to pollute the context window, and also how often it needs to autocompact, and how small the compacted discussion history needs to become not to use too much of the new window.
In that sense, 200k to 1M would be a huge improvement.
Now another question is, will it work linearly better or not? Can "too much context" become a problem - is it still able to understand what part of that context is important and what is not? Frequent autocompact helps here. This is also how many human beings operate, creating notes of key points and studying them.
1. Your coding session can be longer before context compression. I.e. the point when you need to repeat something said earlier happens later.
2. Autonomous code generation can have longer runs. We are now at the level of 2 weeks, and complexity of rewriting OS kernel from scratch.