LangGraph Pricing in 2026: What the Open-Source Library, LangSmith Plans, Deployment and Self-Hosting Cost

LangGraph pricing has two layers. The LangGraph library is MIT-licensed and free to run anywhere. Paid costs come from LangSmith, which since October 2025 also covers LangSmith Deployment, formerly LangGraph Platform: Developer is $0 for one seat and 5,000 base traces a month, and Plus is $39 per seat a month with 10,000 base traces and one free small serverless deployment. Extra traces and deployments are billed in LangChain Standard Units worth $1 each. Self-hosted, hybrid and BYOC deployment are Enterprise-only, at unpublished prices (checked 6 November 2026).

On 1 October 2026 LangChain replaced per-run and uptime charges for deployments with billing on the vCPU and memory a deployment requests, and its price list now meters every product in one unit, the LangChain Standard Unit (LSU). Comparison articles that quote a fee per deployment run describe the retired model. The figures below come from LangChain’s pricing page, its enterprise pricing page, its documentation and the package registries, all read on 6 November 2026. Prices are in US dollars and exclude tax.

What is free and what costs money

LangChain renamed LangGraph Platform to LangSmith Deployment in October 2025, and its documentation now calls the server runtime the Agent Server. A search for “LangGraph Platform pricing” therefore ends on LangSmith’s price list.

LangSmith plans and what each includes

A trace is one execution of your application, however many model and tool calls it contains, so one agent run usually counts as one trace. Base traces stay available for 14 days, extended ones for 180. Since 14 September 2026, 180 days has been the maximum retention for SaaS customers, Enterprise included. Online evaluators and automation rules upgrade the traces they touch to extended retention by default, which raises the bill unless you opt out per evaluator or rule. A seat is any user in the organisation, invited users included.

Other LangSmith products draw on the same LSU meter: Fleet for no-code agents (7 LSUs a month included on Developer and 37 on Plus), Engine, which analyses traces and proposes fixes on a six-hourly schedule (LangChain estimates 7 to 45 LSUs per run, and its enterprise page puts typical use at about 1,000 to 2,500 LSUs a month), Sandboxes for agent-generated code (8 free LSUs a month), Tuned Evaluators (0.015 LSU per evaluation run, in US cloud beta) and the LLM Gateway, whose prepaid model credits carry a 3.5% processing fee. VC-backed startups can apply for up to $10,000 in credits.

Is there a LangGraph Enterprise plan?

Not under that name. LangGraph itself has no paid edition; “LangGraph enterprise” in practice means the LangSmith Enterprise plan, which adds the hosting options, identity controls and support that the library does not have.

Where the data lives and how billing works

On Developer and Plus, LangSmith runs only in LangChain’s cloud, for which the pricing page lists US and EU data locations. Enterprise adds three other arrangements. In a hybrid setup your Agent Servers run in your infrastructure and send traces to LangSmith, either LangChain’s cloud or your own instance. In BYOC, the data plane runs in your own cloud account, connected to LangChain’s control plane, and LangChain operates that cluster for you, upgrades and patching included; it is generally available on AWS with a US control plane and in beta on GCP and Azure. Fully self-hosted means you run everything.

Developer and Plus are billed monthly by card: seats on the first of the month, pro-rated when added mid-month with no credit for removed seats, and traces in arrears. Enterprise is invoiced annually in advance. Workspace spend limits can cap trace costs on the self-serve plans, though LangChain says some of these controls are unavailable on Enterprise because its billing is custom.

LangSmith Deployment pricing, formerly LangGraph Platform

Deployments are billed on the resources each running container requests rather than what it actually uses, sampled every five minutes and totalled hourly. Managed Redis is not billed, and nor is database disk.

LangChain publishes the resulting hourly cost of each size on its own cloud in GCP:

Three details change real bills. A Serverless deployment’s default scale-down time is one week, so one that receives a request at least weekly never scales down; scale to zero is still in beta, and Serverless is not yet offered on AWS or Azure or for self-hosted and hybrid setups. Dedicated deployments bill their minimum replicas and database around the clock, and LangChain recommends them for customer-facing agents. Charges apply per deployment, not per agent, so several graphs defined in one deployment share its replicas and database.

A worked monthly cost example

This example uses only published rates. The volumes and sizing are assumptions, labelled below, and the result is an illustration rather than a quote.

Assumptions: six engineers on Plus; 250,000 agent runs a month, each traced once; 5% of traces (12,500) upgraded to extended retention by an online evaluator; one Dedicated Small production deployment running all month at its minimum size, over the 720-hour month LangChain uses in its own examples; one Serverless Small test deployment inside the free allowance; no Fleet, Engine, Sandboxes or Gateway use; model tokens paid to your model provider.

Trace volume dominates. Doubling runs to 500,000 adds $1,250, while a second agent defined as another graph in the same deployment adds nothing to the deployment line. Each extra Dedicated Small replica that stays up all month adds about $61.56, and a second Dedicated Small deployment adds another $384.12.

Running LangGraph yourself without paid services

You can. Import the library into your own Python or TypeScript service, store checkpoints in your own PostgreSQL with the MIT checkpointer and expose the agent through your own API. Nothing in the MIT licence requires LangSmith. What you give up is everything LangChain sells around the library, which becomes your build:

Two licence boundaries matter. The packaged Agent Server, published on PyPI as langgraph-api, is under the Elastic License 2.0, and LangChain’s CLI reference says langgraph up requires a licence key for production use; standalone Agent Servers read a LANGGRAPH_CLOUD_LICENSE_KEY at start-up. And the free “Self-Hosted Lite” tier, capped at 1 million nodes executed, that LangChain’s earlier LangGraph Platform documentation described no longer appears in its current documentation or price list. If a proposal relies on it, ask for written confirmation.

Air-gapped sites face one more condition. Self-hosted LangSmith verifies its licence and reports usage to beacon.langchain.com. LangChain’s documentation says this billing telemetry cannot be disabled, so installations with no egress need an offline licence from sales.

What drives total cost of ownership

• People. Most of the self-build table is engineering time: a platform team for the runtime and another for approvals, audit and access control. Our breakdown of the hidden costs of assembling a platform from frameworks walks through each category.
• Upgrade cadence. LangGraph shipped six releases between 12 October and 5 November 2026, versions 1.2.15 to 1.2.20 on PyPI, and each upgrade needs a regression run against your graphs.
• Trace volume and retention. At $5 per 1,000 base traces beyond the allowance, observability can cost more than hosting. Sampling helps, and firms that must keep records for years need an export route, because SaaS retention now ends at 180 days.
• Model tokens. Neither LangGraph nor LangSmith Deployment includes model usage; your provider bills it, or your own GPUs serve it.
• Self-hosting infrastructure. Kubernetes, PostgreSQL, Redis, ClickHouse for self-hosted LangSmith, and the staff to patch them.
• Compliance evidence. On the open-source path, approval records, audit trails and agent inventories for the EU AI Act or sector rules are yours to produce. The deployment choices around this are covered in the guide to AI agent deployment models.

LangGraph vs a proprietary on-premises agent platform

This is the comparison many regulated buyers end up making. The table is deliberately even-handed.

LangGraph is the better choice when you have a capable engineering team, an agent design that a visual canvas cannot express, a product that embeds agents for your own customers or a sizeable LangChain investment already. A governed platform is the better choice when many teams build agents under one policy, when approvals and audit must exist from the first release, when nobody wants to operate a runtime, or when sites are air-gapped. Plenty of organisations do both: research in LangGraph, production on a platform.

For a feature-level view, see our side-by-side comparison of VDF AI and LangGraph, the self-hosted LangGraph alternative page and the wider look at LangChain and an enterprise agent platform. If the shortlist includes CrewAI, the orchestration comparison with LangGraph and CrewAI applies, and the framework roundup covers the rest of the field. Architecture questions sit in the companion guides on agentic AI architecture and agent management platforms.

How to estimate your own LangGraph cost

1. Count runs. One execution of your application is one trace. Estimate runs per month for each environment and decide whether development traffic needs tracing at all.
2. Choose retention. Traces live 14 days by default. Decide which share needs 180 days, and check which evaluators and rules upgrade traces automatically.
3. Size deployments. Serverless suits background and internal agents that tolerate cold starts; Dedicated suits customer-facing ones. Group agents into shared deployments where you can.
4. Count seats. Every user in the organisation counts, invited users included.
5. Switch off what you do not use. Engine, Fleet, Sandboxes and Tuned Evaluators all draw LSUs.
6. Add models and infrastructure. Provider token bills or your own GPU costs.
7. Price the people. The engineering in the self-build table is usually the largest line.
8. Ask Enterprise the right questions: the platform subscription price, the price per LSU in your credit commitment, how many instances are included (the enterprise page says one production and one development), self-hosted metering (0.01 LSU per deployment run) and retention terms.

The same method worked for n8n’s plans and licence terms, whose top tier is also priced by quote.

How VDF AI fits

VDF AI is a proprietary agent platform for organisations that want orchestration and governance as a product rather than a build. It runs as VDF AI Cloud, in your own data centre, in a private cloud or VPC, in a sovereign cloud or fully air-gapped. Cloud plans are Free, Starter at $20 a month flat, Professional at $50 per seat a month and Enterprise Cloud through sales. On-premises licences are sold as seat bundles without published prices, and each bundle can begin with a free three-month trial of the platform. Current terms are on the pricing page.

For teams coming from LangGraph, VDF AI Networks is the orchestration layer: a visual canvas with more than 14 node types, including Agent and Multi-Agent nodes, a Human Approval node that pauses consequential work for an accountable person and a Verification node that checks evidence before the next stage. Networks are versioned and validated, retry with backoff, fall back across routes and track cost, latency, tokens and energy for each node. Agents are built in VDF AI Agents with any model per agent and tools from an MCP-powered registry, and Networks, agents and tools can each be called through an API. The migration playbook from LangChain to Networks maps graphs, tools and state onto that model.