A long Claude session doesn't get slow because the work got harder. It gets slow because context piles up without structure, instructions repeated three times, outdated info still sitting in the conversation, a large reference pasted in "just in case" and never removed.

None of that is a model problem. It's a session-hygiene problem, and it's almost always invisible until the session is already bloated and nobody wants to be the one to start over.

What actually helps: separating stable information (things that don't change) from dynamic session state (things that do), closing out finished subtasks instead of leaving their full detail active, and keeping reusable context blocks instead of rebuilding the same instructions every time.

I put together a system for auditing and structuring this properly, checklists, frameworks, and a set of reusable prompt blocks for keeping Claude sessions lean: https://pedronagatomo.gumroad.com/l/ivkdd

If your sessions tend to get messy around the same point every time, it's worth the look.