cache_weight: Explicit Prompt Caching in langchain-openai

What it does

LangChain now lets you explicitly pass cache_weight and related parameters when calling OpenAI chat models, giving direct control over which parts of a conversation are cached. Previously caching was automatic or required manual API-level configuration; now ChatOpenAI accepts explicit caching hints per call without you touching raw OpenAI options.

Why it matters

Prompt caching is one of the most effective ways to cut latency and cost on long conversation chains, especially system prompts that repeat across every request. Before this, you had to build cache_control headers by hand. Now you write ChatOpenAI(model="gpt-4o", cache_weight=0.5) and LangChain translates it into the correct API structure.

Example

$ Initialize ChatOpenAI with a cache_weight hint and invoke the same system plus user message twice.

The function below uses Python's `re` module to validate...\nadditional_kwargs={"‘cached’": True}  # on the second call

Common flags

ChatOpenAI

OpenAI chat model client; now accepts cache_weight

cache_weight

Constructor hint controlling how much input may be cached

BaseChatModel._generate

Internal hook where cache metadata is attached to the response

langchain-fireworks

Sibling package fix reporting cached token usage correctly