Zum Inhalt springen
DeutschlandGPT

Billing

How workflow costs are calculated, tracked, and displayed

Workflow costs are based entirely on actual AI resource consumption. There is no per-run fee, no per-node charge, and no subscription tier that limits the number of workflow executions.

The Workflow Assistant (Chat)

The chat assistant in the workflow builder, the one that builds and edits your workflow from natural-language instructions, is billed pay-as-you-go. It runs on exactly the same basis as your workflow executions: you pay for the AI tokens it actually uses, and nothing else.

Full transparency, by design:

  • You see the cost as it happens. The running total for the current chat is shown directly beneath the chat input, updating after every turn. No estimates, no end-of-month surprises.
  • No hidden or flat fees. There is no per-message charge, no seat fee, and no separate "assistant" subscription. The only thing you are billed for is model usage, charged to the same workspace credit pool as everything else.
  • You stay in control. Pick a lighter model from the selector for small edits and a stronger one only when you need it. Your workspace's spending limits and low-credit alerts apply here too.

Why usage-based and not a fixed price? The assistant works agentically: for a complex request it may inspect your workflow, make a change, run a test, read the result, and refine over several steps, while a small edit might take a single step. A flat fee would over-charge the small requests to subsidise the large ones. Usage-based billing keeps it honest: each request costs what it actually costs. One instruction can stand in for a lot of manual clicking, but you only ever pay for the work the model genuinely did.

Tip: the model selector in the chat lets you trade cost for capability per task. Pick a lighter model for routine edits and a frontier model when you need the assistant to plan or build something complex.

What Costs Money

ResourceBilling Basis
AI Agent nodeTokens consumed by the LLM across all agent steps
Image Generation nodePer generated image (varies by model and resolution)
AI-powered condition evaluationTokens used when a condition uses AI to resolve its expression
AI input resolutionTokens used when a node input uses "AI prompt" mode to generate its value

Free to run (no AI consumption):

  • HTTP Request, Web Search, RSS Feed, Custom Code
  • Condition (when using manual expressions only)
  • Loop, Subworkflow Call
  • All integration actions (Slack, Jira, Gmail, etc.)
  • Data Table operations
  • The trigger itself

In practice, the primary cost driver in most workflows is the Agent node.

How Costs Are Calculated

Each AI call records the number of input tokens and output tokens consumed. The cost is:

CODE
cost = (inputTokens × pricePerInputToken) + (outputTokens × pricePerOutputToken)

Token prices depend on the specific model used. Different models have different per-token rates: more capable models (e.g. Claude Opus, GPT-4o) cost more per token than faster, lighter models (e.g. Claude Haiku, GPT-4o mini).

All costs are denominated in EUR.

Subworkflow Cost Roll-up

When a workflow calls a subworkflow, the costs incurred inside the subworkflow are attributed to the parent execution. The parent run's total cost includes all costs from all nested subworkflow calls. This keeps cost accounting at the top-level execution where it's most useful.

Viewing Costs

Per Execution

Open any execution in the run history to see its total estimated cost. Click into a specific node to see its individual token usage.

Workflow Stats

Open the Stats tab on any workflow to see:

  • Total cost over a selected date range
  • Cost broken down by execution status (completed, failed)
  • Token usage by model: which models were used, how many input and output tokens each consumed, and the resulting cost

Workspace Usage

Go to Workflows → Usage to see an aggregate view across all workflows:

  • Total cost per workflow
  • Trend over time
  • Cost breakdown by model

Budget Controls

Your workspace administrator can configure spending limits and alerts under Settings → Enterprise → General:

ControlDescription
Monthly spending limitCap total workflow AI spend per calendar month. Executions that would exceed the limit are rejected with an error.
Low credit alertReceive a notification when the workspace credit balance falls below a threshold you set.
Auto top-upAutomatically purchase additional credits when the balance drops below a threshold.

Once a monthly spending limit is reached, all workflow executions involving AI nodes will fail until the next calendar month or until the limit is raised. Non-AI nodes in the same workflow are not affected: only the nodes that would consume AI tokens.

Cost Optimization Tips

Choose the right model for the task. A fast, cheaper model (Claude Haiku, GPT-4o mini) is often sufficient for classification, extraction, or routing tasks. Reserve large frontier models for tasks that genuinely require their capabilities.

Limit agent steps. The AI Agent node's Max Steps setting caps how many times the agent can loop. Lower this for simpler tasks to prevent runaway token usage.

Use pinned data during testing. In the test panel, you can pin the output of expensive nodes (like AI agents) so they are not re-executed on repeated test runs. The pinned output is replayed instead.

Use Custom Code for deterministic logic. If a task can be expressed as JavaScript (filtering, sorting, string manipulation, math), use the Custom Code node instead of an AI agent. Code execution is free.

Cache results in Data Tables. If a workflow repeatedly queries the same external data or runs the same AI analysis on the same input, store the result in a Data Table and check the cache first before running the AI call again.

Was this page helpful?