LangChain / LangGraph vs OpenAI Agents SDK
LangChain / LangGraph and OpenAI Agents SDK side by side: what each one is, verified plan pricing and install commands where we have them, and the full review of each.
LangChain / LangGraph
Most popular LLM framework. 100K+ GitHub stars. Chains, RAG, vector stores, tool use. LangGraph adds stateful multi-agent workflows with cycles and persistence.
OpenAI Agents SDK
Python and TypeScript framework for multi-agent workflows. Agent handoffs, tool use, guardrails, sessions, tracing. Provider-agnostic, supporting the OpenAI Responses and Chat Completions APIs plus 100+ other LLMs.
At a glance
| Field | | |
|---|---|---|
| Category | AI Frameworks | AI Frameworks |
| Website | langchain.com | openai.com |
| Pricing | LangChain and LangGraph libraries - free and open source (MIT); LangSmith Developer - $0/seat per month: 1 seat, up to 5k base traces/mo, then pay as you go; LangSmith Plus - $39/seat per month: up to 10k base traces/mo, unlimited seats, 1 free small serverless deployment, then pay as you go; LangSmith Enterprise - custom pricing, self-hosted and hybrid deployment, custom SSO, support SLA; Usage beyond plan allowances is metered in LangChain Standard Units (1 LSU = $1) | Agents SDK - free and open source under MIT (Python and TypeScript); Models - not included; the SDK calls OpenAI models by default, or other providers through the beta Any-LLM and LiteLLM adapters |
| Install | pip install -U langchain or pip install -U langgraph or npm install langchain @langchain/core | pip install openai-agents or uv add openai-agents or npm install @openai/agents zod |
| Models | Model-agnostic: install a provider package (langchain-openai, langchain-anthropic, Google Gemini, AWS Bedrock, Azure, OpenRouter and many more) and pick a model with init_chat_model, for example init_chat_model("google_genai:gemini-3.7-flash"). New model names work without a LangChain update because provider packages pass them straight to the provider's API. | An agent with no model set runs on gpt-5.6-luna, with reasoning effort none and low verbosity. You can override that per agent, per run through RunConfig, or for everything with the OPENAI_DEFAULT_MODEL environment variable. Non-OpenAI models run through the Any-LLM and LiteLLM adapters, which are best-effort betas, and MultiProvider can route on model-name prefixes. |
| Last verified | October 26, 2026 | October 26, 2026 |
In Depth
LangChain / LangGraph
LangChain is the most popular framework for building LLM applications, with over 100K GitHub stars. It provides abstractions for chains (sequential LLM calls), RAG (retrieval-augmented generation with any vector store), tool use, and output parsing. LangGraph extends it with stateful, graph-based workflows - agents that can loop, branch, and persist state across interactions. A more recent addition is 'Deep Agents' for long-running autonomous tasks. LangSmith provides observability and tracing. The ecosystem is massive - integrations with every model provider, vector database, and tool imaginable. I cover LangChain in my AI Agent Frameworks course.
OpenAI Agents SDK
The OpenAI Agents SDK (the production successor to Swarm) is a minimal Python framework for building multi-agent systems. Core concepts: Agents (an LLM + instructions + tools), Handoffs (agents delegating to specialists), Guardrails (input/output validation), and Tracing (built-in observability). It's deliberately lightweight - no heavy abstractions or state management. You define agents as simple Python objects, wire them together with handoffs, and the SDK handles the orchestration. It works with any OpenAI model and supports streaming. Good for building customer service bots, research agents, and workflow automation.
FAQ
Answers taken from each tool's verified fact sheet, with the date it was last checked.
How much does LangChain / LangGraph cost?
As of October 26, 2026: LangChain and LangGraph libraries - free and open source (MIT); LangSmith Developer - $0/seat per month: 1 seat, up to 5k base traces/mo, then pay as you go; LangSmith Plus - $39/seat per month: up to 10k base traces/mo, unlimited seats, 1 free small serverless deployment, then pay as you go; LangSmith Enterprise - custom pricing, self-hosted and hybrid deployment, custom SSO, support SLA; Usage beyond plan allowances is metered in LangChain Standard Units (1 LSU = $1). Source: LangSmith pricing, Install LangChain (Python), Install LangChain (JavaScript), Install LangGraph, LangChain models, langchain 1.4.0 release, langchain-core 1.7.0 release, langgraph releases, langgraphjs releases, LangChain license (MIT).
How do I install LangChain / LangGraph?
Official install commands as of October 26, 2026: pip install -U langchain or pip install -U langgraph or npm install langchain @langchain/core.
Which models does LangChain / LangGraph use?
As of October 26, 2026: Model-agnostic: install a provider package (langchain-openai, langchain-anthropic, Google Gemini, AWS Bedrock, Azure, OpenRouter and many more) and pick a model with init_chat_model, for example init_chat_model("google_genai:gemini-3.7-flash"). New model names work without a LangChain update because provider packages pass them straight to the provider's API.
How much does OpenAI Agents SDK cost?
As of October 26, 2026: Agents SDK - free and open source under MIT (Python and TypeScript); Models - not included; the SDK calls OpenAI models by default, or other providers through the beta Any-LLM and LiteLLM adapters. Source: openai/openai-agents-python README, openai-agents-python releases, Agents SDK models docs, openai/openai-agents-js README.
How do I install OpenAI Agents SDK?
Official install commands as of October 26, 2026: pip install openai-agents or uv add openai-agents or npm install @openai/agents zod.
Which models does OpenAI Agents SDK use?
As of October 26, 2026: An agent with no model set runs on gpt-5.6-luna, with reasoning effort none and low verbosity. You can override that per agent, per run through RunConfig, or for everything with the OPENAI_DEFAULT_MODEL environment variable. Non-OpenAI models run through the Any-LLM and LiteLLM adapters, which are best-effort betas, and MultiProvider can route on model-name prefixes.
