Changing reasoning level results in cache miss

What issue are you seeing?

Within a codex session, changing the effort level to a new value causes a significant cache miss to occur. The order of change does not matter, e.g. low -> medium -> high vs medium -> low. Returning to a previously set reasoning level does not cause a severe cache miss.

The legacy [Prompt Caching 201 (https://developers.openai.com/cookbook/examples/prompt_caching_201#5-troubleshooting-why-you-might-see-lower-caching) guide lists "Changes to reasoning effort" as a potential cause of lower caching (it says, "why you might"). However, that guide states that it applies only to models before GPT-5.6.

The newer Prompt caching guide does not mention anything about reasoning levels and their effect on caching effectiveness.

It is unclear if this behavior is expected or not.

What steps can reproduce the bug?

Begin a Codex session and get a warm cache by sending two short messages. Change to an unused reasoning level, send another message, and observe a cache miss. Change to another unused reasoning level and observe another cache miss. Go back to a previously used reasoning level and observe no regression.

I've attached a simple script below that prints out data from the most recent codex session, for ease of data. Here are two examples showing how a new reasoning level results in cache miss. The second example shows that returning to a prior level does not see a cache invalidation.

The cache miss is essentially a full miss, as the cached value after a reasoning change is the same value from the first turn's cache hit (likely codex's system prompt being cached). Using opencode I observed a 100% miss.

<details>
<summary>low -> medium -> high</summary>

❯ ./tools/codex-session-debug.sh
Session: /Users/danclark/.codex/sessions/2026/07/25/rollout-2026-07-25T19-19-09-019f9b93-4d0b-7de1-8589-13db3be77d82.jsonl
Model:   gpt-5.6-luna

T.C   ACTION           THINKING     INPUT    CACHE     MISS    HIT  D_INPUT  D_CACHE     LOST    OUT    RSN  FLAG
1.1   final response   low          15048     9984     5064  66.3%        -        -        -     17      0  -
2.1   final response   low          15074    14080      994  93.4%      +26    +4096        -     17      0  -
3.1   final response   medium       15101     9984     5117  66.1%      +27    -4096     4096     14      0  CACHE_REGRESSION
4.1   final response   medium       15122    14080     1042  93.1%      +21    +4096        -     10      0  -
5.1   final response   low          15142    14080     1062  93.0%      +20       +0        -      8      0  -
6.1   final response   low          15157    14080     1077  92.9%      +15       +0        -      6      0  -
7.1   final response   high         15171     9984     5187  65.8%      +14    -4096     4096     20     12  CACHE_REGRESSION
9.1   final response   high         15253    14080     1173  92.3%      +82    +4096        -     28      0  -

</details>

<details>
<summary>medium -> low -> medium (no miss) -> xhigh</summary>

❯ ./tools/codex-session-debug.sh
Session: /Users/danclark/.codex/sessions/2026/07/25/rollout-2026-07-25T19-22-09-019f9b96-0ab8-78e3-8f64-63771c2c98b2.jsonl
Model:   gpt-5.6-luna

T.C   ACTION           THINKING     INPUT    CACHE     MISS    HIT  D_INPUT  D_CACHE     LOST    OUT    RSN  FLAG
1.1   final response   medium       15244     9984     5260  65.5%        -        -        -     12      0  -
2.1   final response   medium       15264    15104      160  99.0%      +20    +5120        -      6      0  -
3.1   final response   low          15278     9984     5294  65.3%      +14    -5120     5120     16      0  CACHE_REGRESSION
4.1   final response   low          15302    15104      198  98.7%      +24    +5120        -     11      0  -
5.1   final response   medium       15323    15104      219  98.6%      +21       +0        -      8      0  -
6.1   final response   medium       15339    15104      235  98.5%      +16       +0        -     19      0  -
7.1   final response   xhigh        15368     9984     5384  65.0%      +29    -5120     5120    102     74  CACHE_REGRESSION
8.1   final response   xhigh        15477    15104      373  97.6%     +109    +5120        -      5      0  -

</details>

What is the expected behavior?

I would not expect reasoning changes to affect the cache hit rate. If prompt caching is affected by the reasoning level, explicit documentation would be useful. A warning in the cli would be helpful too.

For example, claude code displays a warning with some newer models that changing effort means “the full message gets re-read on your next message.” With haiku 4.5, that warning is not displayed, and I did not observe a corresponding cache miss.

This suggests that the relationship between reasoning and prompt caching may be model specific. I am not sure if the observed behavior with gpt 5.6 is a bug or undocumented behavior.

Additional information

Script for generating the above
codex-session-debug.sh

Possibly related unexpected cache misses for images: https://github.com/openai/codex/issues/32613