Disable the 1M context window
Claude Code ships with a 1M token context window on the newer models (Sonnet 5, Fable 5, and Opus 4.7 and up). You can turn it off:
CLAUDE_CODE_DISABLE_1M_CONTEXT=1I think you should. We don't need 1M tokens.
The smart zone didn't grow at the same rate when the window went from 200k to 1M. It's around 100-200k. Past that, things get lost in the middle. Chroma shows the same across Claude 4, GPT-4.1, and Gemini 2.5. A bigger window isn't a smarter model. You just get more room to bury the useful bit.
A million tokens every turn is also a great way to burn through your usage limit fast.
Even if Anthropic fixes both tomorrow, I'd still disable it.
This can be difficult to understand at first, because it feels like you're having a conversation with the model. But you're not. It receives the full history every turn and predicts the next token. Every tool call, every stack trace, every dead end you debugged. All of it, every turn. Not just your new message.
A smaller ceiling can act as a signal. When you hit it, you know your context has grown too big. Compact. Develop the discipline to split up your work and be more deliberate about what's in the window.
Constraints can be wonderful things.