LangChain and LangGraph: Versions, Packages, Pricing and Platform Explained

LangChain and LangGraph are two open source Python and JavaScript libraries from LangChain, Inc.: LangChain is an agent framework built around the create_agent harness, and LangGraph is the low-level orchestration runtime underneath it. People often treat them as one product, or confuse both with LangSmith, the company's commercial platform for tracing, evaluation and agent hosting. This page sets out what each library does, which versions are current, how the release and support policy works, what LangSmith costs, where agents can be deployed and how the stack has changed since the 1.0 releases. It is written for developers, architects and the people who approve frameworks for a team. Every fact comes from LangChain's own documentation, repositories, changelog and pricing page. Last checked: 2026-11-15.

This is an independent guide from NHI Mgmt Group, based on LangChain's public documentation as of November 2026. It is not sponsored or reviewed by LangChain. Product features change, so confirm settings against LangChain's current documentation.

Key takeaways

• Two libraries, one runtime. LangChain's create_agent runs on LangGraph, so LangChain is the quick start and LangGraph is where you go for explicit control over state and control flow. LangGraph does not require LangChain.
• Both reached 1.0 on 20 October 2025. Both 1.0 lines are long-term support releases that stay ACTIVE until 2.0 and then get at least one year of MAINTENANCE.
• Legacy support ends in December 2026. LangChain 0.3 and LangGraph 0.4 receive security patches and critical fixes only until December 2026, so teams still on them have weeks, not months.
• The libraries are free; LangSmith is the paid layer. All four core repositories carry the MIT licence. LangSmith has Developer ($0), Plus ($39 per seat per month) and custom Enterprise plans.
• The deployment product has a new name. The documentation now says "LangSmith Deployment is now LangSmith Agent Server", and the docs still redirect older /langgraph-platform/ paths, so all three names appear in search results.

What is LangChain and LangGraph?

LangChain's documentation describes the stack in layers. LangChain is "the agent framework: abstractions and integrations for models, tools, and agent loops". Its central API is create_agent, which the overview page summarises as "Agent = Model + Harness": the harness is the prompt, the tools and any middleware that shapes behaviour around the model loop. LangGraph is "a low-level orchestration framework and runtime for long-running, stateful agents". You define workflows as graphs of shared state, nodes and edges, and can mix deterministic steps with model-driven steps in the same graph. The docs say LangGraph "does not prescribe prompts or agent architecture".

Two further names complete the picture. Deep Agents (deepagents) is an agent harness on top of LangGraph with planning, subagents, filesystem tools and context management, and is still pre-1.0. LangSmith is the commercial platform for tracing, evaluation, prompts and deployment, and it works with agents built in other frameworks too.

LangChain began in late 2022 as a single Python package that Harrison Chase pushed from a personal GitHub account, according to the company's About page, and he co-founded LangChain, Inc. with Ankush Gola in early 2023. The company is headquartered in San Francisco.

Key facts at a glance

As of 15 November 2026, from LangChain's documentation, repositories and pricing page.

Plans and pricing

The open source libraries cost nothing. Money enters through LangSmith, which covers tracing, evaluation, the hosted Agent Server, Fleet (the no-code agent builder) and sandboxes. The table copies langchain.com/pricing as of 15 November 2026. Prices are in US dollars.

How Agent Server usage is metered: the pricing page lists Runtime Compute at 0.0675 LSU per vCPU-hour, Runtime Memory at 0.0090 LSU per GiB-hour, Database Compute at 0.177 LSU per vCPU-hour and Database Memory at 0.025 LSU per GiB-hour, with one LSU equal to $1.00. The LangSmith changelog for early November 2026 adds that BYOC deployments now pay 60% of SaaS rates for Agent Server and database Standard Units.

Availability: deploying to LangSmith Cloud requires a Plus plan or above, and self-hosted Agent Server hosting with a control plane requires the Enterprise plan with Agent Server hosting enabled, according to the deployment documentation. The pricing page does not publish a per-trace overage rate or the extended-retention upgrade fee.

Models available

Neither library ships its own model. LangChain connects to model providers through integration packages, and the provider overview describes "1000+ integrations across chat & embedding models, tools & toolkits, document loaders, vector stores, and more". Many providers have a dedicated langchain-<provider> package. The "popular providers" table on the integrations page, as of 15 November 2026, lists these packages:

Providers are listed alphabetically. The release policy says partner packages maintained by LangChain, such as langchain-openai and langchain-anthropic, follow semantic versioning and are expected to be stable after 1.0, while other partner packages may follow their own policies. Since langchain 1.1 (25 November 2025), chat models expose a .profile attribute with capability data drawn from the open source models.dev project, which middleware and structured output use to decide what a model supports.

Features and modes

LangChain 1.x was rebuilt around three changes listed in the v1 release notes: create_agent as the standard way to build agents (replacing langgraph.prebuilt.create_react_agent), standard content blocks through a content_blocks property that gives unified access to provider features, and a slimmer langchain namespace, with legacy chains and retrievers moved to langchain-classic.

• Middleware: the defining feature of create_agent. Hooks run at before_agent, before_model, wrap_model_call, wrap_tool_call, after_model and after_agent. Built-in examples include PIIMiddleware (redact or block sensitive data), SummarizationMiddleware and HumanInTheLoopMiddleware (approve, edit or reject a tool call).
• Per-call approvals: from langchain 1.4.6 (14 October 2026), HumanInTheLoopMiddleware accepts interrupt_mode="per_call" so each gated tool call raises its own interrupt. LangChain plans to make this the default in the next major release.
• Built-in MCP: langchain 1.4.0 (1 September 2026) added the langchain.mcp namespace, built on FastMCP, replacing the standalone langchain-mcp-adapters package. It is in beta and supports bearer tokens and OAuth 2.1 with dynamic client registration.
• Decisions API: langchain 1.5.0 (11 November 2026) added langchain_core.decisions and two beta middleware: ModelRouterMiddleware, which picks a model per run, and AutoModeMiddleware, which classifies configured tool calls and blocks those it judges risky.

LangGraph focuses on the runtime. Its documented building blocks are durable execution with checkpointing, streaming, human-in-the-loop interrupts and short and long-term memory. Version 1.1 (10 March 2026) added opt-in type-safe streaming and invoke (version="v2"). Version 1.2 (12 May 2026) added DeltaChannel (beta), which stores only the change at each step; per-node timeouts (run_timeout, idle_timeout); node-level error handlers; graceful shutdown through RunControl; and a v3 event streaming API (beta). The release notes say timeouts and error handlers are Python only.

LangSmith Agent Server is described as "a workflow orchestration runtime purpose-built for agent workloads". It is framework-agnostic: the docs cover LangGraph and LangChain apps, Google ADK through the deployments-wrap-sdk package, and other frameworks such as Claude Agent SDK, Strands, CrewAI and AutoGen through the Functional API.

Supported IDEs and platforms

These are libraries, so they run wherever Python or Node.js runs. Development tooling comes from the LangGraph CLI and LangSmith Studio:

• langgraph dev: an in-memory development server, a single Python process that saves checkpoints and memory to a .langgraph_api directory in the working directory.
• langgraph up: builds the project into a Docker image and runs three containers: the API server, PostgreSQL and Redis. Persistent data is stored in PostgreSQL.
• Studio: a graphical interface for an Agent Server, served from smith.langchain.com but run in the browser. The docs state that the data you send to your server is not sent to LangSmith.

Production hosting has four documented options:

Enterprise and admin controls

Admin controls sit in LangSmith rather than in the libraries. As of 15 November 2026 the pricing page lists Organisation Roles (User and Admin) with Google and GitHub sign-in on Developer and Plus, and custom SSO, custom RBAC and ABAC, custom seats and workspaces, a support SLA and annual invoicing on Enterprise. Enterprise is also the plan that offers Hybrid and Self-Hosted deployment, with data in your own VPC for self-hosting. Recent changelog entries show further organisation controls, for example binding workspace service keys to each environment of an external agent, and an organisation-level maximum expiry for service keys. For identity design, roles, service keys and the Agent Server's own authentication, see the LangChain and LangGraph Security Guide and the LangSmith Security Guide.

Data handling and privacy

LangChain's "Data storage and privacy" page sets out what leaves your environment when you run LangGraph locally or in Docker:

The page states that with tracing disabled, no user data is persisted externally unless your graph code calls an external service. When tracing is on, retention follows the LangSmith plan (14 days for base traces, 180 days for extended). What a model provider keeps is governed by that provider: the Deep Agents 0.4 notes, for example, show how to call OpenAI's Responses API with store=False to turn off provider-side storage.

Release history

Dated entries from LangChain's Python changelog and release notes, as of 15 November 2026.

Release and support policy: both projects use semantic versioning. Breaking changes to the public API only arrive in major releases, deprecated features keep working through the whole 1.x series, and deprecated LangChain features still get security updates through all 1.x minor releases. LangGraph says it expects to space major releases at least 6 to 12 months apart, ships minor releases typically every 1 to 2 months, and patches often weekly. LangChain patches can ship "up to a few times per week". APIs marked beta are production-safe but may change; alpha or experimental APIs may change significantly or be removed; anything prefixed with an underscore is internal.

Limits and known constraints

• Python floor moved inside 1.x: langchain 1.5 and langchain-core 1.8 need Python 3.11 or later, while LangGraph still supports 3.10. The release policy allows Python requirement changes for LangGraph only in major releases. Check the interpreter before upgrading.
• Streaming change needs ordered upgrades: the langchain-core 1.7.0 notes say to upgrade every client that reads the stream first (for example @langchain/react 1.2.6 or later, or langchain-core 1.6.13 or later in Python), then the server. To hold the server back, pin langchain-core<1.7.0.
• Some upgrades are one-way: the Deep Agents 0.6.0 notes say rolling back is not supported once threads have persisted, because DeltaChannel writes checkpoints in a new format.
• Beta features: langchain.mcp, DeltaChannel, v3 event streaming and the decision-driven middleware are all labelled beta as of November 2026.
• Community integrations vary: partner packages not maintained by LangChain follow their own versioning and stability policies.
• Language parity: some features land in Python first; LangGraph node timeouts and error handlers are Python only.

Security guidance

This page describes what the libraries are, how they are versioned and what the platform costs. Hardening is covered in the LangChain and LangGraph Security Guide, which goes through credentials, Agent Server authentication, serialisation and logging, and in the LangSmith Security Guide. LangChain's own security policy asks developers to limit permissions to what the application needs, to assume any credential an agent holds may be used in any way those permissions allow, and to combine several layers of defence. Its examples include scoping file system agents to a single directory, giving agents read-only API keys, and issuing read-only database credentials scoped to the tables the agent needs. In framework terms, that maps to HumanInTheLoopMiddleware for sensitive tool calls, PIIMiddleware before data reaches a model, and the MCP tool metadata (such as destructive_hint) for gating destructive tools behind approval. Apply LangChain security updates promptly and report issues through the LangChain Open Source VDP. The wider discipline is covered in the Securing Agent Frameworks Guide.

Practitioner checklist

• Inventory every service still on LangChain 0.3 or LangGraph 0.4 and plan the move to 1.x before support ends in December 2026.
• Replace langgraph.prebuilt.create_react_agent with create_agent, and move legacy chain imports to langchain-classic following the v1 migration guide.
• Confirm Python 3.11 or later before taking langchain 1.5 or langchain-core 1.8, and pin versions in lock files.
• Upgrade stream clients before servers for langchain-core 1.7.0, and test rollback before adopting changes that alter checkpoint formats.
• Decide where agents will run (Cloud, Hybrid, self-hosted or standalone) and check the plan each option needs before committing.
• Set LANGGRAPH_CLI_NO_ANALYTICS=1 and an explicit LANGSMITH_TRACING value in every environment so telemetry and tracing are deliberate.
• Gate sensitive or destructive tools with HumanInTheLoopMiddleware and give agents scoped, read-only credentials where possible.
• Track beta features you depend on (langchain.mcp, DeltaChannel, decision middleware) and subscribe to the changelog RSS feed.

Standards and references

• LangChain: Python changelog
• LangChain: Release policy
• LangChain: Versioning
• LangChain: What's new in LangChain v1
• LangChain: LangGraph overview
• LangChain: LangSmith Agent Server
• LangChain: Data storage and privacy
• LangChain: Pricing

Related NHI Mgmt Group resources: LangChain and LangGraph Security Guide · LangSmith Security Guide · Securing Agent Frameworks Guide · Agent Frameworks vs Workflow Orchestration