openai.internal._rejects_cache_breakpoint

(internal) Whether `model` rejects `prompt_cache_breakpoint`. OpenAI added the field with GPT-5.6; every earlier model answers 400 ("prompt_cache_breakpoint is not supported on this model") on both Chat Completions and Responses, and caches implicitly whatever the request says.

Reference version

Signature

function openai.internal._rejects_cache_breakpoint(model: string) -> bool throws never

(internal) Whether model rejects prompt_cache_breakpoint. OpenAI added the field with GPT-5.6; every earlier model answers 400 ("prompt_cache_breakpoint is not supported on this model") on both Chat Completions and Responses, and caches implicitly whatever the request says.

The list is the OLD families, because that set is closed and the new one is not: a model this does not recognize is treated as one that takes the field, so a new family works without a change here. A model matches a family by being it or by starting with it plus "-", so "gpt-5" covers "gpt-5-mini" and "gpt-5-2025-08-07" but not "gpt-5.6-luna". A fine-tune (ft:<base model>:<org>:<suffix>:<id>) is judged by its base model. Each family is also matched with an "openai." prefix, the model id Bedrock gives an OpenAI model ("openai.gpt-5.5").

chat-latest is a moving alias that points at an old model today.

Source:<builtin>/openai/ns_internal/chat.bamlbytes 3811–4538