The most annoying part of working with Claude on long projects just disappeared. Opus 4.6 now supports a 1M token context window (roughly 2,500 pages of text), and the practical difference is hard to overstate.
Previous context limits meant that extended conversations would trigger compacting, where the model silently summarizes and discards earlier parts of your chat to stay within bounds. You'd get warnings about hitting max chat length. Complex coding sessions, deep research threads, and multi-step writing projects would hit a wall and force you to start fresh, losing all the context you'd carefully built up.
With 1M tokens, that wall is effectively gone for most real-world use cases. You can feed entire codebases, long documents, or maintain marathon work sessions without the model forgetting what you discussed 30 minutes ago. The conversation just keeps going.
What This Actually Changes Day-to-Day
The shift is less about a single big capability and more about removing friction. A few concrete differences:
- Coding sessions can span an entire project without context loss. You can load multiple files, iterate on solutions, and reference earlier decisions without the model losing track.
- Document analysis can handle full-length reports, legal contracts, or research papers in a single conversation rather than chunking them across multiple chats.
- Writing projects maintain consistency across long-form content because the model retains your style notes, outline, and earlier drafts throughout.
The ironic side effect: people are burning through their usage quotas faster. When the experience stops interrupting you, you just keep working. The old context limits acted as an unintentional break timer. Without them, sessions run longer and hit rate limits purely because users are getting more done per sitting.
This is the kind of improvement that doesn't show up in benchmarks but completely changes how the tool feels in practice. Anthropic has been steadily pushing context limits up across the Claude family, but jumping Opus to 1M puts their most capable model in a category where context length stops being something you think about at all.