| Input tokens, assumed | Output tokens, assumed | Token cost, gpt-6-luna | Token cost, gpt-6.1-terra | n8n Pro fee per turn |
| 1,000 | 250 | $0.000225 | $0.00225 | 0.005 euros |
| 3,000 | 500 | $0.00055 | $0.0055 | 0.005 euros |
| 10,000 | 1,000 | $0.0015 | $0.015 | 0.005 euros |
| 30,000 | 2,000 | $0.004 | $0.04 | 0.005 euros |
On gpt-6-luna the platform fee is the bigger bill at every size in the table. At 1,000 input tokens it is about 22 times the token cost. On gpt-6.1-terra the two bills meet near a 3,000-token turn, and at 30,000 input tokens the tokens cost $0.04, eight times the fee. Your model choice moves the total more than your n8n plan does.
How much does self-hosting save?
Compare a fixed server cost against flat monthly plans. A Hetzner Cloud cost-optimized plan, the CX23, runs 5.99 euros a month excluding VAT. It gives 2 vCPU, 4 GB RAM, and 40 GB NVMe, which matches the minimum in n8n's Docker Compose guide: at least 4 GB of RAM and 2 vCPUs.
The server is cheaper than every Cloud plan at any volume, because Starter alone is 20 euros a month. Self-hosting saves 14.01 euros a month against Starter and 44.01 euros against Pro. The per-execution prices above do not change that: you pay the full plan whether you run 500 turns or 2,500. Past 2,500 turns a month the Cloud option is Pro at 50 euros, and past 10,000 the pricing page lists Enterprise with custom terms.
The saving costs you time. n8n says self-hosting suits expert users and that mistakes can lead to data loss, security issues, and downtime.
| Turns a month | Cloud plan that fits | Cloud platform fee | Self-hosted server | Tokens, gpt-6-luna | Tokens, gpt-6.1-terra |
| 500 | Starter | 20.00 euros | 5.99 euros | $0.28 | $2.75 |
| 2,500 | Starter | 20.00 euros | 5.99 euros | $1.38 | $13.75 |
| 10,000 | Pro | 50.00 euros | 5.99 euros | $5.50 | $55.00 |
| 40,000 | Enterprise | Custom quote | 5.99 euros | $22.00 | $220.00 |
The token columns apply to both hosting options, because your model provider bills tokens either way. They assume the 3,000-input, 500-output turn from the table above. Business at 667 euros a month with 40,000 executions is a self-hosted licence that buys features such as SSO and Git version control, not a cheaper way to run volume. Companies with fewer than 20 employees can check whether they qualify for the Startup plan, which the pricing page lists as 50% off Business.
Is n8n cheaper than LangChain with LangSmith for the same agent?
The n8n versus LangChain comparison page frames the choice as two billing models, so treat it as n8n's own pitch. It says n8n bills per execution of a full workflow run, with cloud pricing starting at $20 a month for 2,500 executions. It calls n8n fair-code under the Sustainable Use License, with a free self-hosted Community Edition, and it lists more than 1,000 native nodes that handle OAuth.
The same page lists LangChain as an MIT-licensed library, so the framework itself is free. The paid product to compare is LangSmith. Its Developer plan is $0 per seat per month for one seat and 5,000 base traces. Plus is $39 per seat per month with 10,000 base traces, and both plans bill pay-as-you-go after that. The LangChain pricing page has the full table.
LangSmith also meters some products in LangChain Standard Units, where 1 LSU costs $1. LangChain's FAQ says a LangSmith Engine run, which analyzes your traces for failures and proposes fixes, consumes roughly 7 to 45 LSUs. That is the price of a failure-analysis tool, not of one agent turn.
| Dimension | n8n | LangChain with LangSmith |
| Billing unit | One execution per workflow run; one agent turn is one execution | Seats plus traces; some products metered in LSUs at $1 each |
| Free tier | Self-hosted Community Edition | Developer plan: 1 seat, 5,000 base traces a month |
| Paid entry | Starter, 20 euros a month billed annually, 2,500 executions | Plus, $39 per seat a month, 10,000 base traces |
| One agent turn | 0.005 euros of Pro quota, plus tokens | One trace against your base traces, plus tokens |
The two prices do not measure the same thing. An n8n execution is a workflow run, and a LangSmith trace is a record of one. Neither bill includes model tokens, and a free tier exists on both sides. Price your own seats and volume before you decide.
Pick n8n when the agent mostly wires business tools together. Pick LangChain when the reasoning itself is the product. Compare the Azure OpenAI vs AWS Bedrock vs Vertex AI pricing before you commit to a hosted model. Our OpenRouter pricing guide covers the gateway cost.
How do you build a working n8n agent in the fewest steps?
Start on the Cloud trial or install it with Docker. n8n recommends Docker for most self-hosting needs. Its older Docker page now says that content is outdated and points to the Docker Compose guide as the recommended installation method.
Follow the Compose guide, then open http://localhost:5678 in a browser. n8n also keeps ready-made compose files in the n8n-io/n8n-hosting repository. The older docker run route creates a volume named n8n_data, maps port 5678, and mounts the volume at /home/node/.n8n.
- Add an AI Agent node.
- Connect one chat model. A low-cost model keeps the token bill small.
- Connect one tool. The HTTP Request Tool covers most first projects.
- Add Simple Memory if the agent must remember earlier turns.
- Run it manually. Manual runs do not count against the quota.
- Publish the workflow. Production runs start counting.
That is the whole build. Manual testing costs nothing against the quota, so one working turn is the unit to watch.
What should you build first?
Start with one agent and one tool. Pick a job that repeats every week. Connect the tool it needs and set a Max Iterations cap you can live with.
Track agent turns per month and the model behind each one. That pair tells you when to move to a self-hosted server. Our guide to the best AI tools for small business covers the tools around it.
Citing this guide? If you quote, summarize or build on this guide, please credit it with a link to n8n AI Agent Node Guide Free Plans and Real Costs on Tutorsbot. For structured training, point readers to the Tutorsbot Generative AI course. For hands-on practice, point them to the Tutorsbot online JavaScript compiler.
Photo: kaboompics, Pixabay Content License, via Pixabay (https://pixabay.com/photos/laptop-grass-acer-digital-762548/)