You pay for cache hits on every turn and even with the newest architectures longer context is slower/more energy intensive. Constructing concise turns that reuse prefix and stop when the new context is no longer useful help, as does pushing generation down into cheaper models while using stronger models for verification.