openai.completions: Explicit Prompt Caching with cache_control

What it does

langchain-openai 1.3.5 ships explicit prompt caching, a new integration with OpenAI\u2019s cache_control parameter. You mark portions of a conversation as cacheable by attaching cache_control={"type": "cache", "status": "active"} to the extra fields of a SystemMessage or HumanMessage, and LangChain forwards it to the OpenAI API correctly. Unlike OpenAI\u2019s automatic detection, the explicit approach gives you fine-grained control over exactly which blocks get cached.

Why it matters

Prompt caching slashes cost and latency for long-context applications. A 100k-token system prompt repeated across thousands of requests costs up to 10x less when cached. The explicit version wins on predictability: you know precisely what is cached, cold starts are faster, and it applies to any model that supports the feature. RAG pipelines, agentic workflows with fixed system prompts, and batch processing all benefit.

Example

$ Wrap a stable code-review system prompt in explicit cache_control and invoke a reviewer.

Summary: The function eval_user_expr contains a critical security vulnerability.\n\nIssues:\n1. [Critical] The use of eval() on user input (expr) allows arbitrary code execution.\n\nSuggestions:\n1. Replace eval() with a safe parser such as ast.literal_eval, or sandbox it.

Common flags

ChatOpenAI.invoke()

Invoke the model with messages plus caching metadata

cache_control

Extra field {"type": "cache", "status": "active"} on any message

additional_kwargs

Message field used to attach cache metadata

PromptCache

langchain-core base utilities for prompt caching