The Compact Problem Has a New Enemy

There is a particular texture of dread that sets in around 90% context capacity in Claude Code — you can feel the compaction coming, the way you can feel turbulence before the seatbelt sign comes on.

Compacting is, technically, a feature. It lets you keep going. What it does not do is keep you going well. The quality of anything Claude applies to your in-progress work after a compaction drops, noticeably, in ways that are hard to articulate but easy to feel — like getting a surgery finished by a different surgeon who was handed a two-paragraph summary of where the first one left off.

So the workaround, until today, was manual and faintly absurd: watch your context, aim for a natural stopping point around 5%, scroll back through your terminal, copy maybe 300 lines of history, hit /clear, paste it back in. Fresh context, minimal loss. If the next step was particularly gnarly you could shift-tab into plan mode on the paste and think it through cleanly before touching anything.

This works. It works surprisingly well, actually — Claude doesn't need much to get back up to speed, which is either reassuring or mildly unsettling depending on your mood.

But it requires you to be paying attention. It requires a particular kind of babysitting that is, charitably, not what you opened the terminal to do.

Anthropic just shipped one million context tokens for Sonnet via the API, and Claude Code can use it. The pitch is straightforward: instead of compacting mid-problem, switch to the million-context model, keep everything, finish the thing, wrap up cleanly, then clear and start fresh. Context as a resource you manage deliberately rather than one that manages you.

The catch, because there is always a catch: price does go up past 200k tokens. Not catastrophically — we're not talking about a different order of magnitude — but it's not free. The smart play is still to treat context like working memory and clear it between problems. The million context is a parachute, not a lifestyle.

Whether the promise holds in practice is a different question. A million tokens is a lot of room to accumulate confusion — outdated assumptions, abandoned approaches, contradictory instructions that never got cleaned up. Long context doesn't erase the garbage, it just gives it more space to hide. The manual clear-and-paste discipline had a side effect nobody talks about: it forced you to decide what actually mattered before handing it back.

Still. For that one moment — the tricky bit at the end, the fix that's 80% done, the thing you do not want to hand to a freshly compacted model — this is a real option now.

Will report back.