LangGraph vs CrewAI: which agent framework fits, and what each costs a team of 5 or 25 (October 2026)

LangGraph or CrewAI: the verdict

LangGraph vs CrewAI: comparison table

Compare all fields side by side, including certifications and hosting →

Cost for 5 and 25 people when both libraries are free and the paid layers differ

The two libraries cost $0. Only the hosted platforms around them cost money. For a team of 5 over one year, LangGraph with the LangSmith Plus plan lists at $2,340. For 25 people it lists at $11,700. CrewAI publishes no price above its free plan, so for both team sizes the paid-tier figure reads: Not published by the vendor (as of October 10, 2026).

LangChain's pricing page lists LangSmith Plus at $39 per seat and month, billed monthly and self-serve, and a seat is any user in the organization, invited users included. That gives 5 seats x $39 x 12 = $2,340 and 25 seats x $39 x 12 = $11,700. Plus includes 10,000 base traces a month for the whole organization; each additional trace costs 0.005 LSU, and one LSU is $1, so every extra 10,000 traces add $50. Plus also includes one small serverless deployment; further deployments are metered by vCPU and memory hours. The free Developer plan has one seat and 5,000 traces a month. Nothing forces you to buy any of this to use LangGraph: the documentation says you switch tracing on by setting an API key and can turn it off.

CrewAI's pricing page shows two plans. Basic is free and includes the visual editor, GitHub integration, 2 automations and 50 workflow executions a month, which is 600 a year. Enterprise is listed as custom and adds single sign-on (Microsoft Entra, Okta), role-based access control, deployment in your own VPC and a 45-day onboarding. It shows no seat price, no execution price and no overage rate. For a team of 5 or 25 that wants managed hosting beyond 50 executions a month, the budget therefore starts with a sales call.

Neither figure includes model fees, which you pay to your model provider and which depend on your prompts and run counts, not on the framework. What this means for you: if you only need the library and run it on your own servers, the license cost is $0 either way and the real cost is engineering time. If you want a vendor-managed host, LangGraph gives you a number to put in a budget today, CrewAI gives you a free tier to try and a quote request.

LangGraph draws the process, CrewAI describes the team

LangGraph describes itself as a low-level orchestration framework and runtime for long-running, stateful agents. You define a shared state, add nodes that change it and connect them with edges, including conditional ones. Its package lists langchain-core as a dependency, and the documentation says you do not need LangChain itself to use it. It also exists for JavaScript and TypeScript as LangGraph.js.

CrewAI starts one level higher. A Crew is a team of agents, each with a role, a goal and tools, working on tasks together. Flows are the control layer: methods marked with @start and @listen run when an earlier step finishes, @router picks a branch, and state is a Pydantic model. CrewAI's own documentation describes Flows as the backbone of the application and Crews as the teams a Flow delegates tasks to. Its package file lists no LangChain dependency and requires Python 3.10 up to but not including 3.14.

What this means for you: choose LangGraph if you can already sketch your process as boxes and arrows and want the code to look like that sketch; choose CrewAI if you would rather describe who does what and let the Flow decide when each role runs.

Both save state and wait for a person, but the server rules differ

Durability is no longer a dividing line. LangGraph saves graph state as checkpoints, resumes from them after an error and offers replay and fork from an earlier checkpoint. Its interrupt() function pauses a node and waits until you resume with input. CrewAI Flows have a @persist decorator that keeps state across restarts with SQLite as the default backend, and a @human_feedback decorator, available from version 1.8.0, that pauses a Flow for approval. The differences are in granularity (a checkpoint after each step in LangGraph, a persisted state per Flow method in CrewAI) and in the surrounding tooling.

Hosting is where the two diverge. LangSmith Deployment runs on LangChain's cloud (Plus plan or above), as a hybrid with your own data plane, or self-hosted. Running the Agent Server yourself, outside the cloud, needs a license key, and LangChain's support article says self-hosted LangSmith is available only through the Enterprise plan. The server also contacts LangChain to check the license and count runs unless you run in air-gapped mode. If you embed LangGraph as a library inside your own application, none of this applies. LangChain's documentation also says LangSmith Deployment can host CrewAI agents, so the platform choice is not locked to the framework.

CrewAI's hosted offer is the CrewAI AMP platform, called CrewAI cloud on the pricing page. Its documentation lists crew deployments, REST API access, traces and logs, a tool repository and Crew Studio, a visual editor. The pricing page says Enterprise can run on CrewAI cloud, in your VPC or on your own infrastructure. What this means for you: if you want to host the agents yourself without a license, use LangGraph as a library or CrewAI's open-source package; if you want someone else to run them, compare a priced Plus plan against an unpriced Enterprise conversation.

LangChain publishes more data protection documents, CrewAI names one US cloud

Both frameworks run in your own code, so for the libraries the data stays where you run them. The questions start with the hosted layers. LangChain's trust center lists SOC 2 Type II, ISO 27001:2022, GDPR and HIPAA, with a 2026 SOC 2 Type II attestation. Its subprocessor list names Google Cloud (USA or Netherlands), AWS (USA), Clickhouse (USA or Netherlands) and Supabase (US or Germany). The pricing page says Plus customers choose LangChain's cloud in the US or the EU, and LangChain publishes a data processing addendum (version of October 31, 2025) that references Standard Contractual Clauses.

CrewAI's trust center names SOC 2 as a compliance framework, lists a SOC 2 Type 2 report from June 2026 and a HIPAA audit report from February 2026, and publishes a subprocessor list dated September 21, 2026. That list names Amazon Web Services in the US as the cloud provider. I found no EU hosting region named on the trust center or the pricing page; Enterprise customers can instead deploy into their own VPC or infrastructure. The resources on that trust center include no data processing addendum, so a German or Austrian company should ask for one before real personal data goes in.

Both libraries also send usage data unless you opt out. CrewAI collects anonymous telemetry by default (versions, agent roles, tool names, task counts, no prompts by default); setting CREWAI_DISABLE_TELEMETRY to true turns it off, and the share_crew option adds task details, which can contain personal data. LangGraph's command-line tool logs one analytics event per command, switched off with LANGGRAPH_CLI_NO_ANALYTICS=1, and LangSmith tracing only starts when you configure an API key. What this means for you: set the opt-out variables in your container image on day one, and ask whichever vendor you pick for its data processing terms before you send customer data to a hosted platform.

What users complain about

I looked at public GitHub issue trackers, where users report problems with dates attached. These are individual reports, not vendor-confirmed findings, and I did not reproduce any of them. Status is as of October 10, 2026.

LangGraph: issue 7417 (opened April 5, 2026, still open) reports that tool calls running longer than about three minutes were silently re-dispatched from a checkpoint on LangGraph Cloud, causing 2 to 3 times the work. Issue 3716 (opened March 6, 2025, still open) collects Postgres checkpointer connection errors with an SSL message. Issue 5672 (opened July 25, 2025, still open) says cancelling a streaming run loses state that was not yet saved as a checkpoint. Issue 7714 (opened May 5, 2026, still open) reports that checkpoint serialization adds about 85 percent storage overhead, with no way to switch it off. Cost forecasting is a separate worry: hosted usage is metered in LSUs, so a monthly bill depends on usage you have to estimate.

CrewAI: issue 5802 (opened May 14, 2026, still open) says a task retry runs tools again that already ran, so a payment or email tool can fire twice. Issue 6414 (opened July 1, 2026, still open) asks for a built-in guard against agents delegating a task back and forth or repeating one tool call without end. Issue 4783 (closed March 9, 2026) reported that in the hierarchical process the manager agent could not delegate to other agents. On pricing, the vendor's page lists no number above the free plan. What this means for you: both tools need guards against repeated side effects. Make every tool that sends money or messages idempotent, whichever framework you pick.

What moves, what you rebuild and what you lose when you switch

CrewAI's documentation includes a guide for moving from LangGraph to CrewAI Flows. It maps the StateGraph class to the Flow class, nodes to methods marked @start or @listen, edges to those decorators and @router, TypedDict state to a Pydantic model, and compile plus invoke to kickoff. Its six-step cheat sheet starts with the state model and ends with deciding which nodes should become Crews. The guide is written by CrewAI, so read its claims about boilerplate as a vendor view.

In practice, plan the move in this order. First, your state schema: LangGraph state defined as a TypedDict becomes a Pydantic class, and any reducers that merge state need rewriting. Second, the control flow: conditional edges become @router methods. Third, human approval: interrupt() calls become @human_feedback steps, and approvals still pending have to be created again. Fourth, your stored data: neither vendor documents a converter from LangGraph checkpoints to CrewAI's persisted state, so finish or discard running threads first. Fifth, your traces: history stored in LangSmith stays there, and you need another tracing setup or CrewAI's own.

What is lost depends on direction. Moving to CrewAI, you lose LangGraph.js, since CrewAI's installation page lists Python only, and per-step replay from earlier checkpoints, which I did not find in the CrewAI pages I read; moving to LangGraph, you lose Crews as a ready building block and the roles-and-tasks vocabulary, which you recreate as nodes. Neither vendor publishes a time estimate for a migration (as of October 10, 2026), so I give none. What this means for you: a small project can be rewritten in one sprint of testing, but a system with saved threads and approvals in flight needs a cut-over plan, not just new code.

Neither fits? Look at Haystack

If your application is mostly retrieval and question answering, with a few tool calls and no long-running agent teams, Haystack is the third framework worth reading. It is open source under Apache-2.0, maintained by deepset in Berlin, and its documentation describes pipelines as graphs you wire yourself so they run the same way every time, with agents as an option where you want the model to steer. The repository showed about 26,700 stars on October 10, 2026.

Haystack is not for you if you need a ready-made multi-agent team vocabulary like Crews or a hosted deployment with per-seat pricing published on the vendor site; I did not compare its hosted offer for this page. What this means for you: if both frameworks above feel like more machinery than your use case, a pipeline framework may be the smaller tool.

Who should pick LangGraph and who CrewAI

The bottom line

If you want the process in code, state you can replay and a hosted price you can budget today, LangGraph is the safer starting point. If your work is a team of roles inside a fixed process and a free tier is enough for the first months, CrewAI is the quicker one to try. In both cases, make tools idempotent, switch off telemetry in your image, and get written data processing terms before a hosted platform sees customer data.