Perceived regression in Codex quality and higher token consumption with GPT-5.6

Summary

Since switching to GPT-5.6 in Codex, I have experienced a noticeable decline in day-to-day coding quality together with apparently higher token consumption.

Observed behavior

• The agent more often repeats questions or loses track of information already provided in the thread.
• Small, targeted tasks sometimes trigger broader repository exploration or unrelated refactoring.
• It spends more time explaining or exploring alternatives instead of making a precise minimal patch.
• Cross-file understanding and continuity across turns feel less reliable.
• Fixes can introduce new issues, reducing confidence in allowing the agent to modify code autonomously.
• For comparable tasks, usage appears to consume more tokens than before, without a corresponding improvement in correctness or stability.

Expected behavior

• Preserve relevant context reliably across turns.
• Prefer the smallest necessary change and avoid unrelated edits.
• Limit repository scans and repeated test runs when the task is narrow.
• Use token budget efficiently and provide predictable quality for comparable tasks.
• Make token usage more transparent, ideally broken down by repository reads, reasoning, tool calls, and generated output.

Additional note

This is primarily experiential feedback rather than a controlled benchmark. The key issue is the combination of higher perceived cost and lower confidence in the resulting code. It would be helpful if the team could compare GPT-5.6 against prior Codex models on long-thread continuity, minimal-patch behavior, unnecessary file reads, repeated tool calls, and total token usage per completed task.