LangSmith Deployment is LangChain’s hosting for agents, the product that used to be called LangGraph Platform. You write the agent, in LangGraph or another framework, and its Agent Server runs it with durable checkpoints, resumable streaming, and a task queue. The bill has three parts: seats, traces, and the compute your deployments use.

A Dedicated Small deployment that runs all month comes to about $389 at the published rates, before seats and traces. The Plus plan includes one Serverless Small deployment for free.

Every figure comes from langchain.com/pricing and LangChain’s docs, checked on September 28, 2026. The dollar totals are arithmetic on the published rates, and LangChain’s usage calculator gives the official estimate.

What LangSmith Deployment costs

The plans:

Plan	Price	What it adds
Developer	$0, 1 seat	Up to 5,000 base traces a month, then pay as you go
Plus	$39 a seat a month	Access to Deployment, up to 10,000 base traces a month, 1 free Serverless (Small) deployment
Enterprise	Custom	Self-hosted and hybrid deployment, SSO, RBAC, a support SLA

Deployments bill in LangChain’s own units: an LCU costs $1.50 and an LSU costs $1.00. In dollars, that’s:

What runs	Rate	In dollars
Runtime compute	0.045 LCU per vCPU-hour	$0.0675 per vCPU-hour
Runtime memory	0.006 LCU per GiB-hour	$0.009 per GiB-hour
Database compute	0.177 LSU per vCPU-hour	$0.177 per vCPU-hour
Database memory	0.025 LSU per GiB-hour	$0.025 per GiB-hour

Deployments come in six sizes. Serverless Small, Medium, and Large have 1, 2, and 4 vCPUs and no database of their own. Dedicated Small, Medium, and Large have 3, 5, and 10 vCPUs for the runtime plus a database with 1, 2, and 4 vCPUs.

A month of one deployment

Take a Dedicated Small deployment (3 vCPUs and 6 GiB for the runtime, 1 vCPU and 4 GiB for the database) and run it for a 730-hour month:

Part	Arithmetic	Monthly
Runtime compute	3 vCPU x 730 h x $0.0675	$147.83
Runtime memory	6 GiB x 730 h x $0.009	$39.42
Database compute	1 vCPU x 730 h x $0.177	$129.21
Database memory	4 GiB x 730 h x $0.025	$73.00
Total		about $389

Add $39 per seat and any traces past the plan’s allowance. A Dedicated Medium comes to about $730 the same way. Scale to zero, which would cut idle hours, “is in beta and is initially available only for deployments on the new usage-based pricing.”

What you get for it

• Human-in-the-loop approvals that wait. LangGraph interrupts pause the graph, and it “waits indefinitely until you resume execution.” You resume with the same thread ID.
• Any framework. LangSmith Deployment “is framework-agnostic”: LangGraph natively, and the Claude Agent SDK, Strands, CrewAI, AutoGen, or Google ADK through a wrapper.
• Durable execution. By default, Agent Server writes a checkpoint after each step (its async durability mode), so a thread’s state survives a restart.
• Resumable streams. “Thread streaming also supports resumability: if a connection drops, reconnect with the last event ID to pick up where you left off.” join_stream attaches to a thread or a run that’s already going.
• The US or the EU. Deployments run in either data region, and they can’t move between them afterwards.
• Your agent as an MCP server, on Plus and Enterprise.

What to plan for

1. You still write the agent. Deployment runs your graph. The loop, the prompts, and the tools are your code.
2. Resumed nodes run again. When you resume an interrupt, “the runtime restarts the entire node from the beginning”, so code before the interrupt() call runs twice.
3. Checkpoints pile up. LangChain’s own advice for long conversations is to “Prune old checkpoints periodically or set a retention policy.”
4. Limits. Requests to Cloud deployments are capped at 25 MB, and one thread runs one run at a time.
5. Self-hosted LangSmith is an Enterprise add-on. It needs a license key and brings ClickHouse, PostgreSQL, Redis, and Kubernetes. The lighter standalone Agent Server still needs PostgreSQL and Redis, and LangChain lists the gaps you “must implement and maintain yourself” outside Kubernetes.
6. Trace retention depends on the page you read. The pricing FAQ says extended traces are kept 400 days. The admin docs say 180 days for cloud customers since September 14, 2026.

The alternatives

Where do you go if the seats, the metering, or the self-hosting terms don’t fit? It depends on whether you want to keep writing the agent.

Amazon Bedrock AgentCore

The closest match if you want to keep writing the agent. AgentCore runs any framework in a microVM per session, bills per second from $0.0895 per vCPU-hour with no minimum, and lists HIPAA eligibility, FedRAMP, and SOC 2. The trade is time: sessions last up to 8 hours on microVMs, idle out after 15 minutes by default, and its approval hook waits at most 900 seconds. AgentCore alternatives has the details.

Claude Managed Agents

No agent code to deploy: Anthropic runs the loop and a sandbox per session, and approvals set to always_ask wait indefinitely, like interrupts. You pay $0.08 per session-hour while it runs, plus tokens. The trade is Claude only, one model per session, a beta, and no ZDR or HIPAA.

OpenAI’s Agents API

Also no agent code: OpenAI runs its Codex harness, in beta, with OpenAI models that can change between turns. It’s US-only for data, without ZDR, and it has no approval step of its own.

Cadenya

No agent code, and any model. Cadenya runs the loop and keeps the state: you point a tool set at your OpenAPI document or an MCP server, pick the model on a variation, and start an objective, with approvals set per tool set. It’s free for 1,000 loops a month and $49 for 50,000, where a loop is one LLM request, and there are no seats. What it offers that LangSmith Deployment doesn’t is at the end.

LangGraph on your own servers

LangGraph itself runs anywhere. With a Postgres checkpointer, a graph’s state survives restarts (the in-memory savers don’t: “When the process restarts, all checkpoints are lost”). You give up Agent Server’s queue, scaling, and console, and you run it like any other service.

Side by side

	LangSmith Deployment	Amazon Bedrock AgentCore	Claude Managed Agents	OpenAI Agents API	Cadenya
Who writes the loop	You	You	Anthropic	OpenAI	Cadenya, configured by you
Approval wait	Indefinite	Up to 900 seconds	Indefinite	No approval step	Until answered, within the idle timeout you set
Models	Any your code calls	Any	Claude	OpenAI	Any provider you connect
Regions	US or EU	AWS regions	US workspace geo	US	Not stated
Self-hosting	Enterprise add-on	No	Sandboxes only	Executor only	Nothing to host
Starting price	$39 a seat a month, plus runtime	Per second, no minimum	$0.08 per running hour, plus tokens	Tokens and sandbox time	$0 for 1,000 loops

When to stay on LangSmith Deployment

Stay if your team wants to own the agent’s code, the waits are long, and you need the EU region or a path to self-hosting. That combination is its strength.

What does Cadenya offer that LangSmith Deployment doesn’t?

LangSmith Deployment hosts the graph you write. Cadenya runs the agent itself, so there’s no graph to write, no deployment to size, and no seats to count. Here’s what that covers:

1. No graph to write. You set the agent’s prompt, model, tool sets, and approvals on a variation, and Cadenya runs the loop: it calls the model, runs the tools, streams each reply, and compacts the context window when a conversation gets long.
2. No deployment to size. There’s no Serverless or Dedicated size to pick, no runtime vCPUs or database billed by the hour, and no checkpoints to prune. Your app calls an API.
3. No seats. Plus is $39 a seat a month before any runtime. Cadenya bills per loop, one LLM request: free for 1,000 a month, $49 a month for 50,000, then $0.0015 each. For the $389 a Dedicated Small deployment costs in a month, Growth covers more than 275,000 loops.
4. Your OpenAPI document as the tools. On LangSmith Deployment, tools are your code. A Cadenya tool set turns every operation in an OpenAPI 3 document into a tool, re-checks the document each hour, filters the list, and loads tools on demand. MCP servers and plain HTTP endpoints work too.
5. Approvals that don’t re-run code. A resumed LangGraph node runs again “from the beginning.” A Cadenya tool marked for approval pauses the objective on toolApprovalRequested, and approving runs that one tool call. The answer can come from the API, a webhook handler that posts to Slack, or a person in the React chat widget, and a denial can carry a memo that steers the agent.
6. No second language to run. LangGraph comes in Python and JavaScript, so a Ruby or Go team deploys its agent in a language it doesn’t otherwise use. A Cadenya agent is configuration, and your app talks to it through SDKs for TypeScript on Node 18, Python 3.9, Ruby 3.1, and Go 1.22, or plain HTTP.

How objectives work shows an agent turn on Cadenya, and the free plan needs no card.