This release focuses a lot on cost reductions, and implements "dreaming" among other things, which is rearranging context memory during memory pressure. This probably slashes down your consecutive input tokens by 95% or something, reducing memory pressure, merging rules, removing conflicts, and gets rid of irrelevant text.
It works by as the context reaches 80% of its max amount of messages a "dream prompt" is executed to clean up the context. Inspired by Anthropic's CC leak some 2 months ago.