Summary
I am a ChatGPT Plus subscriber reporting what appears to be a possible GPT-5.6 Sol High reasoning reliability regression.
The issue is not simply that responses are shorter. The issue is that GPT-5.6 Sol High sometimes appears to skip the deeper analysis process that I previously received from the same setting.
Expected behavior
With GPT-5.6 Sol selected and reasoning set to High, I expect complex analytical requests to receive extended reasoning appropriate to the task.
Examples of tasks where this matters:
-
speculative anthropology analysis
-
alternate history modeling
-
fictional worldbuilding consistency analysis
-
multi-factor cause/effect reasoning
-
comparing competing hypotheses
Previously, these kinds of prompts would produce detailed analysis with multiple considerations, tradeoffs, and second-order effects.
Current behavior
On some conversations, GPT-5.6 Sol High now:
-
responds almost immediately even to complex analytical prompts;
-
produces answers that feel more like summaries than analysis;
-
does not consistently explore alternatives or failure points;
-
misses deeper implications that would normally be expected;
-
feels closer to a fast-response mode despite High reasoning being selected.
This is not fixed by:
-
starting a new conversation;
-
regenerating the response;
-
increasing reasoning settings (High is already the maximum available on my Plus account).
Reproduction example
Prompt:
“Analyze whether a technologically advanced but psychologically traumatized species could build a stable interstellar civilization. Approach it like a speculative anthropology paper. Compare at least three models and identify failure points.”
Without additional instructions, GPT-5.6 Sol High produced a much more generic overview than expected.
However, adding:
“Before answering, build an internal framework: define assumptions, identify variables, create competing hypotheses, test them against each other, then write the final analysis.”
produced a noticeably better result.
This suggests the model is capable of deeper analysis, but may not be consistently deciding to enter that mode by default.
Comparison test
GPT-5.5 High appears to behave closer to my previous expectations when given the same analytical framework.
The difference seems to be less about capability and more about default reasoning behavior.
Possible causes (hypotheses only)
I cannot determine the backend cause, but possible explanations include:
-
reasoning effort allocation not being applied consistently;
-
complexity detection underestimating analytical prompts;
-
routing/session state issues;
-
changes to default response optimization;
-
a regression in GPT-5.6 Sol’s reasoning behavior.
I am not claiming that a specific cause is confirmed.
Why this matters
For users who use ChatGPT primarily for factual lookups, shorter answers may not be a problem.
For users who use it as an analytical tool, the difference is significant. The value is not just the final answer, but the ability to model a problem, compare possibilities, and explore consequences.
Request
Could OpenAI confirm whether there are any known issues with GPT-5.6 Sol High failing to consistently apply extended reasoning on complex prompts?
If this is intended behavior, clarification on how High reasoning is supposed to differ from previous behavior would also be appreciated.
I can provide example conversations showing the difference between expected and current behavior.