Suddenly seeing nearly 2x as many GPT-5.6-Luna turns as Sol while only using Sol

What version of the Codex App are you using (From “About Codex” dialog)?

26.810.52044

What subscription do you have?

plus

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

I have consistently used gpt-5.6-sol in Codex Desktop and have never used gpt-5.6-luna.

Starting on Aug 15, 2026, Codex Analytics suddenly began showing nearly twice as many Luna turns as Sol turns. For example, on Aug 16, when the spike was largest:

• gpt-5.6-sol: 124 turns
• gpt-5.6-luna: 232 turns

Before this date, Analytics only showed a small amount of occasional Luna usage.

What steps can reproduce the bug?

1. Use Codex Desktop normally with gpt-5.6-sol selected.
2. Open Codex Analytics and view Turns by model.

What is the expected behavior?

I would like to understand what happened. Why did Codex Desktop suddenly generate so many Luna turns when I was only using Sol?

Additional information

I checked the local Codex telemetry to verify whether these Luna turns were merely an Analytics display issue.

These Luna sessions consistently used approximately the following configuration:

model = gpt-5.6-luna
reasoning_effort = Low
approval_policy = Never
sandbox_policy = ReadOnly

They did perform real Responses inference and generated token usage.

For Aug 16, the local telemetry contains:

• 241 unique ephemeral Luna thread IDs
• 232 completed Luna turns with token usage
• 2,219,106 total Luna tokens
• about 9,770 tokens per completed Luna turn on average

Codex Analytics also reports exactly 232 Luna turns on Aug 16.

These Luna threads are not normal persisted threads, and there are no corresponding rollout JSONL files. All persisted sessions are Sol, which matches the model I intentionally used.

No source files or tests are identified. Start by comparing Codex Analytics' Turns by model with the local telemetry for the ephemeral Luna thread IDs, then trace how those sessions are created and counted; done means identifying why Luna inference occurred and whether the analytics totals are correct.