Summary

This operating policy says the 1M context window is real but should be used deliberately. The right model is tiered context management, not blanket max-window usage.

Key points

• "Supports 1M" is not the same as "best at 1M."
• Accuracy and economics both get worse as context expands.
• 272K is a practical guardrail: pricing cliff, quality inflection, and quota multiplier zone.
• Large context is best for orientation and mapping, not precision changes.
• The strongest pattern is a two-pass workflow: wide orientation, then narrow execution.

Implications

This is directly useful for agentic coding and research work. It provides a sane context policy for when to go wide and when to go tight.

Open questions

• Which current workflows should get an explicit two-pass context policy?
• Should context-tier choice become part of task-card defaults?

Sources

• /Users/kalenhowellsr/Projects/corp-strategy-work/2. areas/ai-dev-practices/assessments/2026-03-07-gpt54-1m-context-window-operating-policy.md