To spot it, reconcile both invoices against the same run count. If model spend per run dwarfs platform spend per run, a cheaper plan saves you almost nothing. ### How loops and retries inflate LangGraph costs Every time it fires, a node inside a loop bills again, not once per graph. A reflection step that critiques and rewrites, or a tool call that retries on failure, multiplies executions on exactly the runs that already go badly. To spot it, compare nodes drawn on your graph with nodes executed per run in your traces. A wide gap means your estimate is wrong. Set a hard iteration cap before launch. If the agent mostly moves data between business apps, the metering unit is the thing to change. Activepieces charges 1 credit per flow run however many steps the flow has, so a tenth step costs the same as a second, whereas Zapier's pricing page bills per task and Make's bills per module. MoneyGram, Moneypenny, Alan and FundingSocieties run Activepieces in production.  ### Retention, seats and enterprise-only features Keeping traces costs money, and longer retention or higher volume moves you into overage. Seats add a per-person charge that grows with the team, and some capabilities sit only on enterprise tiers. Those often need a sales call to price. To spot it, list the features you need and confirm in writing which tier includes each before you commit. ## LangGraph Platform against its closest alternatives on cost The five options differ less in entry price than in what they meter and what you operate yourself, and those two columns decide the monthly bill. The table shows that structure, with entry prices left unconfirmed rather than guessed. | Option | Pricing model | Entry price | Metered items | What you operate | Check date | |---|---|---|---|---|---| | LangGraph Platform | Plan plus usage | Not confirmed | Node executions, deployment uptime, traces | Your graph code | Not verified | | Self-hosted LangGraph | Free library, you pay for infrastructure | No licence fee for the library | Your own compute, database, monitoring | Everything | Not verified | | AWS Bedrock AgentCore | Consumption | Not confirmed | Runtime compute and memory time, plus related services | Agent code and its configuration | Not verified | | CrewAI | Hosted tiers | Not confirmed | Not confirmed | Agent definitions | Not verified | | Mastra Cloud | Hosted tiers | Not confirmed | Not confirmed | Agent and workflow code | Not verified | | Activepieces | Plan tiers (Free, Plus, Team, Ultimate, Embed) plus pay-as-you-go credits over the allowance | $0 (Free plan, 1,000 credits a month, unlimited flows) | Credits: 1 credit per run however many steps it has; AI steps cost 2, 10 or 20 credits by model tier; $0.007 a credit over the allowance on Plus and Team | Your flows and agents on Activepieces Cloud, or your own servers if self-hosted | 2026-10-23 | Every unconfirmed cell is a question for the vendor's own pricing page. ### Self-hosting open-source LangGraph Because the library carries no licence fee, the bill moves to servers, a database for checkpoints, and the engineer who is paged when it falls over. That makes the cost fixed and visible, but it only beats a managed plan if your team's time is cheap. ### Managed agent runtimes from other vendors AgentCore bills by consumption, which rewards bursty agents and punishes ones that sit idle holding state. CrewAI and Mastra Cloud sell hosted tiers, so read what each tier caps before assuming it covers your volume. ### Which pricing model suits which workload Steady, high-volume graphs favour fixed infrastructure you control. Spiky or low-volume agents favour consumption pricing. If your graphs run many nodes per run, look hardest at per-node metering, since that's where the total grows fastest. ## Fix agent cost sprawl with Activepieces ### When a workflow replaces the hosted graph If your agent mostly reads and writes business apps and calls a language model at only a few steps, an Activepieces workflow can replace the hosted graph, and the per-node arithmetic mostly stops mattering: Plus is $20 a month flat for 10,000 credits, and a run is one credit however many steps it has.  Activepieces has a core under the MIT licence, set out in the LICENSE file of its GitHub repository, with only the `packages/ee/` directories under a separate commercial licence, and it is self-hostable, so you can run it on infrastructure you already pay for and see the bill before you commit. ### What the workflow looks like The screenshot shows the shape of that fit. A five-step workflow runs from a webhook trigger, to a database read, to an "Ask ChatGPT" step, to an SMS send, to a database write.  The selected step holds a question field with support-agent instructions and a business-details block describing StoreX, an Apple repair shop open 8am to 10pm Pacific. Only one of the five steps calls a model, so model latency becomes your limiting factor. The other four are plumbing, and plumbing's the part you least want metered like reasoning. ### What it costs to run Pricing it is easy: one model call per run, plus whatever your host charges for the machine. There's no graph to keep alive between requests and no tracing layer to pay extra for. The business context lives in a text field you can read and edit, so a support-bot change is a text edit rather than a redeploy. ### Where it stops fitting The fit has limits. Long-running agents that checkpoint state, pause for a human, and resume mid-graph are what LangGraph is built for, and Activepieces's 746 integrations don't change that; a workflow builder won't reproduce it. The two also combine well. Keep LangGraph for the reasoning loop, and let Activepieces handle the tool layer: the database reads, the SMS sends, the app writes, split across as many small flows as you like, since every plan, including free, allows unlimited flows. That keeps the metered graph small, and the small graph is the cheaper one. ## A Monday checklist for pricing a LangGraph deployment Price a deployment by multiplying nodes per run by monthly runs, then adding deployment type, token cost and tracing, in the order below. The sum is the bill; the plan name is not. 1. Open one real trace and count the nodes executed in a single run, retries and loops included, because the metered unit follows what ran rather than what you drew. 2. Estimate monthly runs from actual traffic, not launch hopes, since this multiplier moves the total more than any plan choice. 3. Pick development or production deployment, because an always-on production deployment bills for uptime even when nobody calls it. 4. Add LLM token cost per run, which often outweighs the platform charge and won't appear on the LangGraph invoice. 5. Reread the pricing page and write down the date you read it, so a later surprise can be traced to a change rather than a mistake. 6. Set spend alerts on usage and tracing before launch; a runaway loop costs the same as a busy day. 7. Decide whether to self-host by pricing your own engineering hours against the managed uptime charge. 8. Set an upgrade rule in writing, such as moving up only when measured usage crosses a threshold you chose in advance. | Option | What you pay for | What to measure first | |---|---|---| | Development deployment | Testing capacity and usage | Nodes per test run | | Production deployment | Uptime plus metered runs | Monthly runs and idle hours | | Self-hosted | Your own infrastructure and staff time | Hours spent on operations | The name changed, so older comparisons describe an earlier price sheet. Treat any figure quoted under the old name as stale until you have matched it against the current page and noted the date, as in step five. ## Frequently asked questions Most of the questions readers send trace back to the old-name problem, so these answers assume you are checking each one against the current pricing page. ### Is LangGraph Platform free to try? There's an entry plan you can start on without a card, but it caps usage, so it tells you whether the agent runs, not what it costs. Price a realistic node count and uptime before you trust a free trial as a forecast. ### Is the open-source LangGraph library free for commercial use? Yes. The library is released under a permissive open-source license, so you can ship it in a paid product. What costs money is the hosted deployment, runs and tracing around it, not the graph code. ### Is LangGraph Platform now called LangSmith Deployment? Yes, the hosted deployment product was renamed and folded under the LangSmith brand. That's why older blog posts and comparison tables quote figures that no longer map cleanly to current plan names. ### Can I run LangGraph Platform in my own cloud? Self-hosted and bring-your-own-cloud options exist, but they sit on higher tiers, and you still pay for the infrastructure underneath. Ask in writing what the license covers. If the answer requires a sales call, treat the price as unknown until you have a number. ### How do I stop a runaway agent from running up my bill? Cap it at three layers: - Set a recursion limit on the graph so a loop ends instead of cycling. - Put a hard spend limit and alert on your model provider account, since token charges sit outside the platform bill. - Sample your traces rather than recording every run, because tracing volume is metered too. A loop that burns through nodes all night costs the same whether anyone was watching or not. ## Related reading - [Snowflake Pricing Explained for 2026 Data Teams](https://www.activepieces.com/blog/snowflake-pricing-explained-for-2026-data-teams) - [Best AI Agent Platform in 2026: Top 6 Tools Compared](https://www.activepieces.com/blog/top-6-ai-agent-platforms) - [n8n AI Agent Node Explained: Tools, Memory, Cost](https://www.activepieces.com/blog/n8n-ai-agent-node-explained-tools-memory-cost) ## References - [LangChain pricing](https://www.langchain.com/pricing) - [Bex](https://bex.co/blog/2026/09/19/langgraph-platform-ga-agent-runtime-vs-git-push-paas)Cost per run is nodes executed times the sum of node fee and model spend, and the plan covers neither term.