claude-code - 💡(How to fix) Fix [FEATURE] Document optimal context window ranges for 1M token models [1 participants]

Summary

With the move to 1M token context windows, users observe significant quality degradation well before the theoretical limit. It would be very helpful if Anthropic documented the optimal operating ranges.

Observed behavior

In extensive multi-session testing with Opus 4.6 (1M context) on a complex project (~55 wiki pages, multi-agent coordination):

• 200K-400K tokens: Performance is strong, reliable recall, accurate task tracking
• ~500K tokens: Noticeable degradation — agent forgets items from only a few interactions ago
• ~50% context (500K): In one session, the agent forgot the session's primary objective entirely. The /loose-ends scan (which shares the same degraded context) also missed it.

This suggests training is optimized for ~100K tokens (the middle of the previous 200K window), and the 1M extension stretches beyond the trained sweet spot.

Request

1. Document the effective operating ranges — at what context percentage should users expect quality degradation? Is there a recommended "plan to finish by X% of context" guideline?
2. Consider documenting this in Claude Code's help/docs — users planning long sessions need to know that 1M tokens ≠ 1M tokens of effective work.

Why this matters

Long sessions are common for complex tasks (architecture review, multi-file refactoring, wiki maintenance). Users invest significant time building context, only to have quality silently degrade. Knowing the effective range in advance would let users plan sessions around it rather than discover it through failure.