codex - 💡(How to fix) Fix Regression: Token usage significantly increased in v0.121.0 vs v0.116.0 for identical prompt and unchanged workspace [5 comments, 4 participants]

What issue are you seeing?

After upgrading OpenAI Codex from v0.116.0 to v0.121.0, token usage increased significantly (by ~60%+) for the exact same prompt, executed in the same workspace, with no code or configuration changes.

Interestingly, v0.116.0 appears to be the most efficient version, while both v0.115.0 and v0.121.0 use substantially more tokens. This suggests a regression in context handling introduced after v0.116.0.

Environment Codex versions tested: v0.115.0 v0.116.0 v0.121.0 Model: gpt-5.3-codex medium OS: Windows Workspace: same project, no changes between runs Prompt: identical in all cases

What steps can reproduce the bug?

Open the same project workspace Run Codex with version v0.115.0

Ask: Please summary me the code within the workspace Note token usage Repeat with: v0.116.0 v0.121.0 Ensure: no changes to code no changes to configuration same prompt used

What is the expected behavior?

Token usage should remain relatively stable across versions for the same prompt and workspace, or change only slightly due to minor improvements.

Additional information

Actual Behavior

Token usage varies significantly between versions:

v0.115.0
total=63,395
input=60,290 (+570,368 cached)
output=3,105
reasoning=813

v0.116.0 ✅ (most efficient)
total=52,896
input=50,392 (+333,056 cached)
output=2,504
reasoning=548

v0.121.0 ❌ (regression)
total=86,194
input=83,231 (+673,664 cached)
output=2,963
reasoning=1,040

Summary

• v0.116.0 → best efficiency (lowest input, cached, and reasoning)
• v0.115.0 → moderate usage
• v0.121.0 → clear regression (highest across all metrics)