LangGraph is LangChain’s low-level orchestration runtime for stateful AI agents. Where LangChain gives you higher-level agent APIs, LangGraph gives developers explicit control over graphs, state, persistence, interrupts, streaming, and multi-agent workflows.
System Verdict
Pick LangGraph if you are building production agents that need more than one LLM call. It is strongest when the workflow must branch, loop, pause for approval, persist state, recover after failure, or coordinate multiple agents.
Skip it if the job is a simple prompt, tool call, or no-code workflow. Direct provider SDKs, LangChain’s higher-level agent API, Mastra, CrewAI, or LangFlow can be faster depending on the team and abstraction level.
Commercial reality: the LangGraph library is free and MIT-licensed. The paid layer is LangSmith: Developer is $0/seat/month (5k base traces), Plus is $39/seat/month (10k base traces plus one dev-sized deployment), Enterprise is custom, and LangSmith Deployment adds $0.005 per deployment run, $0.0007/minute dev uptime, and $0.0036/minute production uptime on top of any per-trace and Fleet costs.
Key Facts
| Last verified | June 12, 2026 |
| License | MIT for the library |
| Languages | Python and JavaScript/TypeScript |
| Current product shape | LangGraph for orchestration; LangSmith for tracing, evaluation, prompts, and deployment |
| Official proof points | LangChain’s docs/site list companies including Klarna, Uber, J.P. Morgan, Replit, Elastic, and more |
| Core concepts | Durable execution, persistence, streaming, interrupts, memory, subgraphs |
| Library cost | $0 |
| LangSmith plans | Developer $0/seat/month (5k base traces); Plus $39/seat/month (10k base traces, one dev deployment, 500 Fleet runs); Enterprise custom |
| LangSmith traces | $2.50 per 1k base traces (14-day retention); $5 per 1k extended traces (400-day retention) |
| LangSmith Deployment | $0.005 per deployment run; $0.0007/minute dev uptime; $0.0036/minute production uptime; $0.05 per additional Fleet run |
| Hosted deployment naming | LangSmith Deployment, formerly LangGraph Platform |
| Related | Deep Agents: LangChain’s higher-level harness built on LangGraph |
When to pick LangGraph
- Long-running stateful agents. Research pipelines, support workflows, code agents, and process automations that need to remember state and resume cleanly fit LangGraph’s design.
- Durable execution matters. LangGraph is built for workflows that can recover from failures and keep running over time, not just one request-response loop.
- Human-in-the-loop workflows. Interrupts let a team inspect, edit, approve, or reject state before the agent continues.
- Multi-agent or subgraph systems. Subgraphs help split complex agent behavior into smaller units without giving up central state control.
- LangSmith is already part of the stack. The first-party tracing, evals, Studio, and Deployment path are smoother when the team is comfortable with LangChain’s ecosystem.
When to pick something else
- Simple one-shot LLM calls: direct OpenAI, Anthropic, or Google SDKs.
- TypeScript-first agents with product-style ergonomics: Mastra is purpose-built for TypeScript teams.
- Visual no-code or low-code building: LangFlow gives a canvas for LangChain and LangGraph workflows.
- Role-based multi-agent prototypes: CrewAI is easier when the mental model is “crew of specialist agents” rather than explicit graph state.
- Azure-aligned enterprise: Microsoft Agent Framework is the more natural fit when Azure AI Foundry is the default platform.
What changed recently
- June 8 source recheck: LangGraph’s current docs still position it as the low-level orchestration runtime for durable execution, streaming, human-in-the-loop control, memory, and persistence, with LangSmith as the tracing/evaluation/prompt/deployment platform.
- LangSmith deployment costs remain itemized. Public pricing still breaks out $0.0007/minute for development deployment uptime, $0.0036/minute for production deployment uptime, and $0.005 per additional deployment run.
- Trace pricing is still split base vs extended. Base traces and extended traces remain separate meters, with base retention at 14 days and extended retention at 400 days by default.
- Deployment scope widened in the docs. LangSmith Deployment is now framed as framework-agnostic agent infrastructure: LangGraph/LangChain are first-party paths, while Deep Agents, Google ADK, and other frameworks can also route through the same Agent Server runtime.
- Fleet and infrastructure meters matter. Plus includes 500 Fleet runs/month, then $0.05 per additional Fleet run; current pricing also lists LangSmith Engine LCUs and sandbox compute as separate meters.
Pricing
| Product | Current public price | Notes |
|---|---|---|
| LangGraph library | $0 | MIT-licensed open-source framework |
| LangSmith Developer | $0/seat/month | One seat plus 5k included base traces per month |
| LangSmith Plus | $39/seat/month | 10k included base traces, one dev-sized LangSmith Deployment, 500 Fleet runs per month |
| LangSmith traces (base) | $2.50 per 1k | 14-day retention |
| LangSmith traces (extended) | $5 per 1k | 400-day retention; $2.50 per 1k to upgrade base traces into extended retention |
| LangSmith Deployment runs | $0.005 per run | One end-to-end invocation of a deployed LangGraph agent |
| LangSmith Deployment uptime | $0.0007/minute dev; $0.0036/minute production | Listed per minute the deployment database is live |
| LangSmith Fleet runs | $0.05 per additional run | Plus includes 500 Fleet runs per month |
| LangSmith Enterprise | Custom | Adds alternative hosting, custom SSO/RBAC, and dedicated support |
Pricing verified June 12, 2026 against LangChain’s public pricing and LangSmith billing docs. Model/API provider costs are separate from LangSmith.
Failure Modes
- Low-level by design. LangGraph does not hide the agent architecture. Teams must still design state, tools, routing, evals, and deployment behavior.
- Billing has multiple meters. Seats, base traces, extended traces, deployment runs, dev uptime, production uptime, Fleet runs, and third-party model costs can all matter once LangSmith is in production.
- Idle dev deployments still bill. A development LangSmith Deployment that sits idle still incurs $0.0007/minute of uptime (~$30/month per dev deployment); production deployments are roughly five times that. Tear down unused deployments.
- Model spend is separate. LangSmith can price seats, traces, deployments, Fleet, Engine, and sandbox compute, but OpenAI, Anthropic, Google, Mistral, or local-infrastructure model costs still sit outside the LangSmith bill.
- Verbose for simple work. call or straightforward tool call is usually cleaner with a direct SDK.
- LangChain conventions still show up. LangGraph can be used standalone, but many real projects use LangChain message, tool, and model abstractions around it.
- Python examples are the deepest path. JavaScript/TypeScript support is real, but Python remains the more mature documentation and community path.
Against the Alternatives
| LangGraph | Mastra | Microsoft Agent Framework | CrewAI | AG2 | |
|---|---|---|---|---|---|
| Primary fit | Stateful production agents | TypeScript agents and workflows | Azure enterprise agents | Role-based crews | AutoGen-style multi-agent work |
| Main language | Python + JavaScript/TypeScript | TypeScript | .NET + Python | Python | Python |
| License posture | MIT library | Open source | Open source | Open source | Open source |
| Control model | Explicit graph state and edges | Agent/workflow primitives | Agents plus workflows | Agents with roles/goals | Conversational agent patterns |
| Hosted path | LangSmith Deployment | Mastra Cloud/platform path | Azure AI Foundry | CrewAI Enterprise | Self-host/community first |
| Best for | Durable stateful agents | TS-first product teams | Microsoft/Azure buyers | Fast crew prototypes | AutoGen continuation |
Methodology
Produced by the aipedia.wiki editorial pipeline. Last verified June 12, 2026 against LangGraph official materials, LangGraph Python and JavaScript docs, LangChain pricing, LangSmith billing docs, LangSmith Deployment docs, the LangGraph GitHub repository, and the LangChain/LangGraph v1.0 announcement.
FAQ
Is LangGraph really free? Yes. The LangGraph library is MIT-licensed and free to use. Paid spend starts when a team uses LangSmith services such as observability, evaluation, or hosted deployment.
What replaced LangGraph Platform? LangChain’s current public product language calls the hosted layer LangSmith Deployment. LangSmith billing docs describe it as the deployment product formerly called LangGraph Platform.
Do I need LangChain to use LangGraph? No. LangGraph can be used standalone, but the docs commonly use LangChain components for models and tools, and many teams pair it with LangSmith for tracing and evals.
What’s the difference between LangGraph and LangChain? LangChain is the higher-level agent framework. LangGraph is the orchestration runtime underneath or beside it when you need durable execution, persistence, human-in-the-loop, streaming, and explicit state control.
Does LangGraph compete with Mastra? Yes, especially for TypeScript teams choosing an agent/workflow framework. LangGraph has the broader LangChain/LangSmith ecosystem; Mastra often feels more native for TypeScript-first product engineering.
Sources
Review History
- 2026-06-12: Re-verified LangGraph/LangSmith pricing and deployment docs, removed unsupported Claude model/vendor release claims, added current LangSmith Engine, sandbox, Fleet, deployment, and framework-agnostic deployment context.
- 2026-05-13: Itemized LangSmith Deployment uptime ($0.0007/min dev, $0.0036/min production), base vs extended trace pricing, Fleet runs, and added a price_history block.
- 2026-05-10: Refreshed LangGraph pricing, hosted deployment naming, source-backed buyer guidance, and current verification language.
- 2026-04-18: Initial agent-framework review.
Related
- Category: AI Automation and AI Coding
- Compare: LangGraph vs Mastra - LangGraph vs CrewAI
- See also: LangFlow - Microsoft Agent Framework - Langfuse
