The 1M token context window lets Claude Code hold roughly a full mid-size codebase plus history in a single session - no compaction, no chunking.

What it does

Supported Opus and Sonnet models on qualifying plans offer up to a one-million-token context window. That's enough for a large repository's source, a long conversation history, and tool output without losing earlier turns. Compaction still exists as a safety net but rarely fires during normal work.

When to use it

Large codebases you'd otherwise have to chunk.
Long debugging sessions with lots of file reads.
Tasks that require tying together evidence from many files at once.
Agent runs where compaction loss would hurt correctness.

Gotchas

More context costs more per turn, even with caching. Watch your spend on marathon sessions.
Availability depends on your plan and the selected model. Check /status if you're unsure.
Just because the window is big doesn't mean you should fill it. Focused reads still outperform broad ones.

Official docs: https://code.claude.com/docs/en/model-config.md#extended-context