What does the title refer to? Have a look at this screenshot from the API reference, reflecting the public specification:
It says gpt-5. Then gives reasoning.context. Is there any reason here to assume that all reasoning models on the API wouldn’t give you control of persisting or dropping reasoning turns via “all_turns”? That.. this would a message input handling method of the API instead of a model-based option?
Guess what. :
gpt-5.4: OK
gpt-5.2: Error: Unsupported value: ‘all_turns’ is not supported with the ‘gpt-5.2-2025-12-11’ model. Supported values are: ‘auto’ and ‘current_turn’.
Also nowhere: if this parameter will damage self-managed reasoning replayed if not set to “all_turns”, or if it is only for server-side chat products and not a large “input”.
mode:“pro” in the API documentation doesn’t mention only gpt-5.6:
Maybe it could be a model ID switcher for old models to go -pro too? You’d not know by reading the API documentation that it is only gpt-5.6 and the removed pro model ID slug.
Then you have new parameters such as prompt_cache_options, setting explicit mode (the net effect is only to disable an automatic cache write that happens at the final user turn of input). This is a change for gpt-5.6, where that top-level parameter is noted as for gpt-5.6.
But then look at the schema of messages of input and the accessory “prompt_cache_breakpoint”:
This could easily be ignored and tolerated if not supported. Nothing to tell you which models are rejecting “prompt_cache_breakpoint”. Send what the API reference documents, gpt-5.5 will error-out with support for only one “24h” type write.
OpenAI could even backport the single undocumented cache write 5.5 gives, and make it respond to “explicit” - so finally gpt-5.5 could be serviceable for task and batch reuse.
Not to mention the bizarre, such as constantly-changing default reasoning.effort per new model in the gpt-5 family, or reasoning.effort “minimal” on gpt-5 but no aliasing “none” to that, or that because of defaults, gpt-5.4 can tolerate temperature with no reasoning sent but not gpt-5.5 (now that more reasoning models with reasoning.effort:“none” will accept temperature and top_p, along with undocumented Responses additions of “frequency_penalty” and “presence_penalty”.)
Conclusion
API Reference needs concrete grounding of the model cutoff of offered parameters, and models pages need a minimum of at least their default reasoning.effort and accepted reasoning levels.
Ingredients keep getting thrown in to a single endpoint’s validation soup; at least a “beta” need was recognized for internal agentic mode.