An n8n workflow’s cost comes from the paid APIs it calls, not from n8n. Every AI call bills for the words going in and coming out, every web search adds a fee, and search results get billed as input words too. One of my workflows made 206 AI calls with web search and cost about $63 a run. Input tokens alone were $52.20 of it. Now Claude warns me before anything over $5 a run goes live.
The rate-limit screen I didn’t understand
On October 1 I opened my Claude Console and saw a rate-limit screen I didn’t understand, so I asked Claude what it meant. One minute that morning had used 96% of my input limit. Claude traced it back to one workflow.
It was the quarterly refresh for my subscription benefits database, the one that keeps track of perks for things like Walmart+, Prime and Costco. Every three months it looked up the perks on the web to see what had changed. That run made 206 calls to Claude Sonnet 4.5, and each call could run up to 8 web searches. It added up to 17.4 million input tokens and 794 searches, about $63 for one run. It had been running every quarter since July 1, and I had no idea.
Holy crap. I need to be warned when a workflow is so expensive.
I turned it off that day. The next morning a $75 spending alert showed up, and about $69 of it came from that refresh and a monthly price check that ran in the same minute.
How much does an n8n workflow cost to run?
An n8n workflow costs whatever its paid services charge per run, multiplied by how often it runs. n8n itself doesn’t bill per AI call. The AI model, web search, image and transcript services do. A small AI workflow can cost under a penny a run. A big one with web search can cost dollars.
API costs add up, and a big, beefy workflow can cost a lot more than you’d expect. Here’s where my $63 went, using Anthropic’s published Sonnet 4.5 prices:
| Piece of the run | How it’s billed | Cost |
|---|---|---|
| 17.4 million input tokens | $3 per million | $52.20 |
| 794 web searches | $10 per 1,000 | $7.94 |
| Everything else, mostly the written answers | $15 per million output tokens | about $3 |
| Total | about $63 |
The search fees were the small part. The big part was input, and the reason surprised me. Anthropic’s pricing page says web search results count as input tokens. Every search pulled pages of results into the call, so each of the 206 calls was reading about 84,000 tokens before it wrote a word.
The rough sizes, so you have some expectations going in. All prices come from each company’s own pricing page, checked October 22, 2026:
| What you’re paying for | Price | What one costs |
|---|---|---|
| Claude Haiku 5.5, input | $0.10 per million tokens | A 1,000-token prompt is $0.0001 |
| Claude Sonnet 5.5, input / output | $2 / $10 per million tokens | A 1,000-token prompt with a 500-token answer is about $0.007 |
| Claude Opus 5.5, input / output | $4 / $20 per million tokens | The same call is about $0.014 |
| OpenAI gpt-6-luna, input / output | $0.10 / $0.50 per million tokens | The same call is under $0.001 |
| OpenAI gpt-6-astra, input / output | $10 / $50 per million tokens | The same call is about $0.035 |
| Web search, Claude or OpenAI | $10 per 1,000 searches, plus the results as tokens | One search is $0.01 before the results |
| AI image (Grok, Nano Banana, OpenAI) | $0.006 to $0.24 per image, depending on model, quality and size | One image can cost more than 10 short AI calls |
| Video transcript (Supadata) | 1 credit per transcript, about $0.006 a credit on the Pro plan | 100 transcripts is about $0.57 |
| Web page scrape (Firecrawl) | 1 credit per page | Plans start at 5,000 credits for $19 a month |
The jump from a short call to a search-heavy call is the one to watch. A short Sonnet 5.5 call is less than a penny. A call reading 84,000 tokens of search results is about 17 cents on the same model. Run that 206 times and you’re at $35 before a single search fee.
These are estimates from published prices on the date I checked them. Your real bill can come out different, since prices change and retries add calls. Check your provider’s billing page, and set a spending limit there too. I’m not responsible for what a workflow ends up charging you.
Who needs an n8n cost check?
You need one if any of your workflows call a paid AI model in a loop, use web search, generate images, or run on a schedule you don’t watch. Those are the ones that run up a bill without telling you.
Skip it if any of these sound like you:
- Your workflows don’t call paid APIs. Moving data between apps, sending emails and updating a spreadsheet cost nothing extra per run on self-hosted n8n.
- You only run one or two AI calls at a time, by hand. You’ll see the cost in your provider’s dashboard long before it matters.
- You’re on a flat subscription, not API credits. A chat plan doesn’t bill per token, so there’s nothing per run to estimate.
What do you need to check n8n workflow costs before they run?
- Your provider’s pricing page. Prices change, so the check reads the current page instead of trusting numbers saved in a file.
- A rough count for each workflow. How many AI calls per run, about how big each one is, how many searches, images or transcripts, and how often it runs.
- Claude Code or any AI chat. I build with Claude Code, and the rule lives in its memory. The generator below also writes a paste-in prompt for other chats.
- A dollar limit. Mine is $5 a run. Pick your own.
How does the cost warning work?
It’s a rule Claude follows, not a workflow. Before any workflow that calls a paid API gets built or changed, Claude estimates the cost per run from calls times tokens times price, plus any search, image or transcript fees. It says the dollar figure in plain words before anything goes live. Search results count as input in the estimate, since that’s what caught me.
If a workflow comes in over $5 a run, Claude flags it in its own short message and offers a cheaper way to build it:
- Fetch the page directly when you already know the address. Fetching a known page has no search fee, and it pulls in one page instead of a pile of results.
- Lower the number of searches per call. Mine allowed 8 per call.
- Use a cheaper model. The same 17.4 million input tokens cost $1.74 on Haiku 5.5 instead of $52.20 on Sonnet 4.5. Haiku 5.5’s price is for prompts up to 100,000 tokens, and mine averaged about 84,000.
- Use the Batch API. Anthropic takes 50% off input and output tokens for batch jobs that don’t need an answer right away.
- Move it off the API. This is what I did.
The quarterly refresh still runs, just not on API credits. Since October 14, n8n drops a task into my work queue on the 5th of January, April, July and October, and Claude does the check on my regular Claude subscription. The $63 version stays off.
The warning hasn’t come up yet. Nothing I’ve built since has crossed $5 a run. I’d rather have it sitting there and never fire than find out from a spending alert again. My n8n monitoring dashboard shows what each workflow cost after it runs. This is the check before it runs.
Get the generator
My rule is built around my workflows and my $5 limit. The generator asks about yours instead. Describe one workflow: how many AI calls per run, which model, about how big each call is, how many searches, images or transcripts, and how often it runs. You’ll see the estimated cost per run and per month, and which piece costs the most.
Then set your own limit, and it writes you a skill for Claude Code, a prompt you can paste into any chat, or both. It’s the same check as mine, set to your limit and the services you use.
The skill file goes in a .claude/skills folder. The prompt gets pasted into Claude, ChatGPT or whichever chat you use. The same disclaimer applies here: these are estimates from prices checked October 22, 2026, and your real bill can differ.
What it doesn’t do
It doesn’t watch your bill. It’s a check at build time. If a workflow starts looping or retrying after it ships, the estimate won’t catch it. A spending limit in your provider’s dashboard and an n8n error workflow are what catch those.
It doesn’t know exact token counts ahead of time. Search results especially vary by what comes back. The estimate gets you the right ballpark, not the penny.
It can’t fix old workflows on its own. It fires when a workflow gets built or edited. The $63 one was built before the rule existed, and nothing looked at it again until the rate-limit screen did.
It doesn’t cover n8n’s own plan. If you’re on n8n Cloud, that’s a separate subscription. Mine is self-hosted, so n8n itself costs me nothing per run.
Frequently asked questions
Does n8n charge per AI call?
No. Self-hosted n8n doesn’t charge per call at all. The AI provider bills you for tokens, searches and images through your own API key, and n8n Cloud charges for its plan separately.
Why is web search so expensive in an AI workflow?
The search fee is small, about $10 per 1,000 searches on both Claude and OpenAI. The bigger cost is the results. Anthropic counts web search results as input tokens, so every search fills the call with pages of text you pay to send.
How much does one Claude API call cost?
A short call on Sonnet 5.5, with a 1,000-token prompt and a 500-token answer, is about $0.007. A call reading 84,000 tokens of search results is about $0.17 on the same model. The size of what goes in matters more than the size of what comes out.
How do I estimate an n8n workflow’s cost before it runs?
Multiply calls per run by tokens per call by the model’s price, then add search, image and transcript fees and multiply by runs per month. Count search results as input. The generator in this post does the math for you.
What’s the cheapest way to cut an expensive AI workflow?
Switching models usually saves the most. My 17.4 million input tokens would cost $52.20 on Sonnet 4.5 and $1.74 on Haiku 5.5. Fetching known pages instead of searching, capping searches per call and the Batch API’s 50% discount help too.
Should I set a spending limit on my AI API account?
Yes. A $75 spending alert the morning after is what put a number on mine. Set a limit or alert in your provider’s console before you turn on anything that runs on a schedule.
Try it yourself
An n8n workflow’s cost is calls times tokens times price, plus search, image and transcript fees, times how often it runs, and search results count as input. Start with the generator above and your biggest scheduled workflow. If the estimate surprises you, check its last run against your provider’s usage page before it runs again.
A tool I built
Subscription Graveyard
Bury the subscriptions you’ve canceled, give each one a tombstone, and watch what you’re saving pile up.
Bury your first one →I run an affiliate program on the tools I build. Approved affiliates earn 20% on what their referrals pay, for up to a year. US only for now. See the program →
Related reading
- My n8n Monitoring Dashboard Gives Every Workflow a Heartbeat
- My Software Bill Is $583 a Month. The Blog Is $328.
- Before You Build Another n8n Workflow, Check the ROI
- Claude Code Memory: How I Stop Repeating Myself Every Day
Subscribe if you haven’t yet, and check one scheduled workflow this week before it checks your bank account for you.
