OpenAI lowered Codex CLI’s default GPT-5.6 input window to 272,000 tokens from 372,000, shrinking how much code, chat history and session state the agent retains before compacting older context.
That smaller memory buffer has drawn complaints from developers on Reddit and X, who say long-running sessions and large repositories could hit compaction sooner and require more resets or manual context handling.
Analysts said routine bug fixes may see little change, but repository-wide refactors and enterprise-scale workflows could suffer from repeated searches, lost earlier decisions and engineers spending more time supervising the tool.
The change also highlights a broader enterprise risk: managed AI coding platforms can alter context limits, pricing and runtime behavior with little notice, pushing teams to build retrieval-heavy, modular workflows that tolerate shifting vendor constraints.