n8n OpenAI Cost Tracking: Know Which Workflow Spent the Money

A $400 OpenAI Bill and No Idea Which Workflow Caused It

If you run more than one AI workflow, n8n OpenAI cost tracking is the difference between "the bill went up" and "the resume-screening workflow tripled its token use after last week's prompt change." The OpenAI dashboard can't tell you that. It reports one total per key and stays silent on which workflow, which model, or which run drove the spend.

There's no strong page ranking for this exact problem, which is odd, because it's one of the most common complaints from people running LLMs in production. The fix is straightforward: tag every call, log the tokens and cost, total it daily, and alert before the budget breaks.

n8n is the right place to do this because the spend originates here. Every OpenAI node knows its own workflow, its model, and its token count. You just have to capture it.

What You Can Automate in Cost Tracking

A spend-tracking layer turns an opaque bill into a daily report:

• Per-call logging: capture tokens, model, and workflow name on every OpenAI call
• Cost calculation: convert token counts to dollars using current per-model pricing
• Per-workflow attribution: see which automation spends what, not just the grand total
• Daily rollup: a scheduled summary of yesterday's spend by workflow and model
• Budget alerts: fire a warning when daily or monthly spend crosses a threshold
• Circuit breaker: route calls to a cheaper model or stop them when the cap is hit
• Trend log: a Google Sheets history you can chart to spot a creeping cost regression

The Cost Tracking Pipeline

[any workflow] → OpenAI node (returns usage: prompt/completion tokens)
  → Code (cost = tokens × per-model price; attach workflowName, model)
  → Google Sheets (append: date, workflow, model, tokens, cost)

Schedule Trigger (daily 23:30)
  → Google Sheets (read today's rows)
  → Code (sum cost by workflow + grand total)
  → IF (total > dailyBudget?) → Slack + Gmail (budget alert + top spender)
  → Slack (daily digest: spend by workflow)

1. Read token usage off the OpenAI node

The OpenAI node returns a usage object alongside the completion. Pull the counts straight from the response:

// Code node right after the OpenAI node
const u = $json.usage ?? {};
const promptTokens = u.prompt_tokens ?? 0;
const completionTokens = u.completion_tokens ?? 0;

// per-1K-token prices for the model you're calling (update to current pricing)
const inPer1k = 0.0025, outPer1k = 0.01;
const cost = (promptTokens / 1000) * inPer1k + (completionTokens / 1000) * outPer1k;

return [{ json: {
  date: new Date().toISOString().slice(0, 10),
  workflow: $workflow.name,
  model: $json.model ?? 'gpt-4o',
  promptTokens, completionTokens,
  cost: Number(cost.toFixed(4)),
} }];

$workflow.name is the attribution the OpenAI dashboard can't give you. That one field is what turns a total into a breakdown.

2. Log every call

Append the cost row to Google Sheets. One row per call is fine; sheets handle tens of thousands of rows without complaint, and you'll roll them up daily anyway. The columns that matter: date, workflow, model, tokens, cost. With those you can answer almost any spend question after the fact.

3. Roll it up on a schedule

A separate Schedule Trigger workflow reads the day's rows near midnight, sums cost by workflow, and posts a digest. This is where the value shows up: a daily message that says "AI spend yesterday: $18.40, of which resume-screening was $12.90" makes a cost regression obvious the day it starts, not the day the bill arrives.

4. Alert before the budget breaks

An IF node compares the running total against your threshold and fires when it crosses. Name the top spender in the alert so the message is actionable, not just "you're over budget." In practice a daily cap catches runaway spend faster than a monthly one, because a prompt change that doubles token use shows up within 24 hours instead of three weeks later.

5. Add a circuit breaker for the expensive workflows

For the workflows that can spike, go further than alerting. Before the OpenAI call, read today's running total. If it's over the cap, route to a cheaper model or a No Op that skips the call. This trades some quality for a hard spending ceiling, which is the right trade for a background batch job even if it isn't for a customer-facing one.

Implementation Patterns

Pattern 1 — Tag at the source. Capture $workflow.name on every call, right after the OpenAI node, before the data flows anywhere else. Attribution added later is guesswork. The AI Agent Spend Tracker template wires this capture into a reusable sub-workflow so you tag once and reuse everywhere.

OpenAI → Code (cost + $workflow.name) → Google Sheets

Pattern 2 — Daily cap beats monthly cap. A daily threshold surfaces a cost regression inside a day. A monthly one lets it run for weeks. Set both, but make the daily one the alert you actually watch.

Schedule (daily) → sum today → IF (> dailyCap) → alert with top spender

Pattern 3 — Circuit-break the batch jobs. For non-interactive workflows, read the running total before the call and downgrade the model or skip when over budget. Keep customer-facing calls on the alert-only path.

n8n Nodes You'll Use Most

Getting Started

1. After each OpenAI node, add a Code node that reads usage and computes cost.
2. Tag the row with $workflow.name and the model.
3. Append every call to Google Sheets.
4. Build a second workflow with a Schedule Trigger to roll up daily spend.
5. Add an IF node and a budget threshold that fires a Slack and email alert.
6. For batch workflows, add a pre-call circuit breaker that downgrades or skips when over the cap.
7. Chart the trend monthly, then start from a template instead of rebuilding the cost math per workflow.

Cost tracking is the financial mirror of error tracking; the same logging-and-alert shape powers the n8n log alerting workflow. And if you're building the OpenAI calls these costs come from, the parsing and prompt tips in n8n OpenAI workflows keep token use predictable.