n8n Cost Tracking: What Every Workflow & LLM Node Really Costs

n8n cost tracking means attributing what each workflow, execution, and LLM node actually costs — so "this workflow ran 4,000 times" becomes "this workflow spent $312 on GPT-4o last month, and a third of it was retries." n8n's execution log won't tell you that: it records whether a run succeeded and how long each node took, but not the provider, model, tokens, or dollars behind your AI nodes. This guide shows how to see n8n cost per workflow, per execution, and per token — and where the waste usually hides.

Why n8n doesn't show you cost

n8n was built to run automations, not to account for them. The execution list answers one question — did every node finish without throwing? — and nothing about money. LLM nodes (OpenAI, Anthropic, the AI Agent node) do return token usage in their output, but n8n never aggregates it. There's no per-workflow spend view, no per-model breakdown, and no alert when a workflow's cost suddenly triples.

For a handful of workflows you can live with that. Once you're running dozens, or once real budget flows through your LLM nodes, the invoice from OpenAI or Anthropic arrives as one lump sum with no idea which of your 40 workflows caused it. That's the gap n8n cost tracking closes.

The three numbers you actually want

Good n8n cost tracking gives you spend at three levels, and each answers a different question:

How to track n8n cost — three ways

First, the DIY route: every LLM node exposes token usage in its output (usually `usage.prompt_tokens` and `usage.completion_tokens`). Add a node at the end of each workflow that reads those, multiplies by the model's per-token price, and appends a row to Google Sheets or a database. It works, but you maintain the pricing table by hand, it breaks when a node changes, and you still have to build the dashboard.

Second, community workflow templates: several n8n templates capture token cost into a DataTable or Sheet. They're a fine starting point, but they're per-instance, need manual setup, and give you a flat table rather than trends, alerts, or cross-workflow rollups.

Third, connect an observability tool read-only. Obsivara ingests an `llm_call` event per model call — provider, model, prompt/completion tokens, cost, latency — via the generic webhook, the SDK, or native n8n, with no proxy in your request path. From that it attributes spend per workflow, per execution, per model, and per token automatically, flags waste and retries, and alerts on cost anomalies. The complete guide to monitoring n8n workflows covers the full connect steps; if you just want a number first, the free LLM cost estimator models your monthly spend (including the retry tax) in a minute.

Where the money actually hides

Once cost is attributed per workflow, the leaks are obvious. Retries and silent failures are the biggest: a tool call times out, the agent retries, and you pay for every attempt — at a 15% failure rate that's real money producing nothing (the same failures covered in debugging silent failures in n8n). Prompt bloat is next: system prompts, retrieved context, and chat history grow over time, so the average call costs more this month than last. And model drift — a workflow quietly switched to a bigger model, or a new node added — steps cost up with no single obvious cause.

Track cost per workflow week over week and you're looking for two shapes: step changes (a config change or new consumer) and steady creep (growing context or retries). Token-per-execution is the leading indicator — it moves before the invoice does. For the full playbook, see the true cost of unmonitored AI.

See your real n8n spend

You can build this yourself with the n8n API and a time-series store, or connect a tool that does it out of the box. Obsivara syncs your n8n executions read-only — no logging nodes to maintain, no proxy in the path — attributes cost per workflow, model, and node, and alerts when spend drifts. Start with a free 48-hour audit to see what your workflows actually cost, or create a free account.