[Bug] Switching from Luna Max to GPT-6-astra can sometimes cause abnormally high token consumption

What version of Codex CLI is running?

codex-cli 0.154.0

What subscription do you have?

business standard

Which model were you using?

gpt-5.6-luna-max

What platform is your computer?

Microsoft Windows NT 10.0.19045.0 x64

What terminal emulator and version are you using (if applicable)?

PowerShell 7.6.6

Codex doctor report

What issue are you seeing?

When switching models within an existing Codex session, I have noticed inconsistent token consumption after switching from Luna Max to GPT-6.

I understand that GPT-6 normally consumes tokens faster than Luna Max, so some increase is expected.

However, there appear to be two very different behaviors:

In some sessions, switching from Luna Max to GPT-6 results in the expected increase in token consumption.
In other sessions, after the exact same type of model switch, token consumption becomes abnormally fast, far beyond the normal GPT-6 usage rate I observe in other sessions.

The abnormal behavior appears intermittently. It does not happen every time the model is switched.

This makes it look less like the normal difference in token cost between the two models and more like something related to model switching, context handling, prompt-cache reuse, compaction, or usage accounting.

What steps can reproduce the bug?

1. Start or continue a Codex session using Luna Max.
2. Work normally for some time.
3. Switch the existing session from Luna Max to GPT-6.
4. Continue working with GPT-6.
5. Observe token / usage consumption.
6. Repeat this across multiple sessions.

Sometimes GPT-6 usage is normal after the switch.

Other times, usage starts decreasing significantly faster than normal despite a comparable workload.

What is the expected behavior?

Switching from Luna Max to GPT-6 should result in reasonably consistent token consumption for comparable workloads.

GPT-6 may naturally consume more tokens than Luna Max, but switching models should not occasionally cause a very large unexplained increase in consumption.

Additional information

There are several existing issues related to token consumption and model switching, including:

#19877 — discusses prompt-cache continuity when switching models in an existing session.
#44462 — reports unexpectedly high context/token growth with GPT-6.
#43222 — reports disproportionately fast GPT-6 usage/quota depletion.

However, the behavior I am reporting is specifically about the intermittent nature of abnormal consumption after switching models: the same Luna Max → GPT-6 transition can sometimes behave normally and sometimes consume tokens dramatically faster.

It may be useful to investigate whether model switching can occasionally cause:

loss of prompt-cache reuse,
duplicated/replayed context,
unexpected context expansion,
additional compaction/reconstruction requests,
stale model-specific session state,
or incorrect usage accounting.

If there is local telemetry that would help diagnose this, I can provide session/thread logs from both a normal switch and an abnormal switch for comparison.

No source files or tests are identified. Review issues #19877, #44462, and #43222, then compare session or thread logs from normal and abnormal Luna Max to GPT-6 switches; done means identifying the cause of the intermittent excess consumption or documenting the missing evidence needed to reproduce it.