1M

Two days ago I wrote about ordering a 128GB Mac. Yesterday Anthropic announced the max subscription for Claude Code now supports a 1M context window. 200K to 1M. The timing was kind of funny.

Same ratio as going from 24GB to 128GB of RAM. Same feeling too.

I’ve been testing it for a full day. The biggest change is compaction. Before, Claude Code would compress the conversation pretty often. Context gets long, the agent summarizes everything, you lose details. (Every time Claude Code asked me how it performed on compaction I pressed “1”: bad.) Sometimes it forgets what files it read or what it already tried. I was used to it but it was always there.

With 1M the compaction still happens, just much later than I’m used to. A few times I was sure it was about to compact and it just kept going. Kind of unfamiliar. I keep bracing for something that doesn’t come.

The other thing I noticed is the stability. When the context doesn’t get compressed as often, the agent holds a more consistent picture of the whole task. Doesn’t re-read files, doesn’t forget what it tried. The work gets better because the agent can hold everything in its head at once.

SKILLS.md is still a good practice for organizing what gets loaded into context. People engineered these to control what’s relevant per task, and it was seen as a really creative piece of agentic structure engineering. How to make agents effective within a token budget. I wrote about SKILLS.md a few days ago and said this form might live forever. I still think that. But the reason shifts from saving space to organizing information.

Lance Martin shared a note about this. Anthropic removed the pricing surcharge past 200K, flat pricing to 1M now. His point was that apps like Claude Code succeed by aligning with model improvements, not building around older limitations.

A lot of context engineering patterns people built for 200K were really workarounds for limited space. With 1M those constraints are loosening.

Context is really just RAM. When computers didn’t have enough they swapped to disk and everything was slow. More RAM didn’t make the CPU faster, it just let the machine hold more state, and that changed what you could build. The engineering people spent on carefully managing agent context might just need one prompt now. A lot of old best practices for agent structure are becoming unnecessary. Better to dig deeper into other parts of the problem than keep holding onto solutions that the models are outgrowing.

8GB of RAM used to be plenty. Now it’s not enough for anything serious. 200K context is heading the same way.