Back to guides
Pricing·September 21, 2026·10 min read

n8n AI API cost: one execution, many model calls, two bills

Separate the one-execution platform bill from the per-call model bill, then find the n8n node defaults that quietly multiply the second one.

Last reviewed on .

n8n bills you once per workflow run; your AI model API bills you for every model call inside that run. Those are two separate meters, and only the first one has a quota. n8n's own definition leaves no room: "An execution is a single run of your entire workflow. It doesn't matter how many steps are in the workflow or how much data it processes" (n8n pricing page, checked September 19, 2026). An AI Agent node sitting inside that one execution can make up to ten model round-trips by default, and the OpenAI Chat Model node retries a failed call twice — so a single execution worth 1/2,500th of a Starter plan's quota can carry up to roughly thirty billed model calls.

One disambiguation before the numbers, because the query splits three ways. The n8n public REST API — the one that lists and triggers workflows — carries no published per-call price; what n8n's documentation states about it is that it is not available during the free trial and asks you to upgrade (n8n API docs). An n8n plan is priced by executions and is covered below purely as the contrasting bill. The model tokens your workflows burn is the subject of this page. n8n is active and unrenamed: this page reads its behaviour from the release n8n@2.39.8, published September 18, 2026, and newer tags have shipped since (GitHub releases).

Meter one: what an n8n execution counts

Execution accounting is the reason the two bills move independently. A hundred-node workflow and a two-node workflow each cost one execution. What varies is which runs count at all.

Run typeCounts toward quota?Rule as documented
Schedule TriggerYesOne execution every time the node fires, regardless of outcome
Webhook TriggerYesOne execution per inbound request that activates the trigger, including an empty body
Polling triggerOnly when data is foundPolls that return no data do not count
Manual run from the editorNoManual executions are excluded
Sub-workflow via Execute Sub-workflowNoOnly the parent top-level execution counts
Error workflowNoError workflow executions are excluded

All six rules are quoted from n8n's executions documentation, checked September 19, 2026. One thing that page does not settle, so it is not asserted here: whether a manual retry of a failed execution counts. It does settle the neighbouring case — a schedule trigger counts "regardless of outcome", and a malformed webhook request that fails before the workflow starts does not count. Note also that quotas are not a Cloud-only concept — the same page states that paid plans, cloud or self-hosted, have an execution limit quota.

The sizing example n8n publishes in the FAQ on its pricing page is useful: a daily schedule produces 30 or 31 executions a month, while one running every five minutes produces about 8,600 to 8,900. The same FAQ states that saved-execution history is capped separately — by count, by storage on Cloud, and by a retention window — and that reaching one of those limits does not stop workflows from running, it only shortens the history kept. Worth keeping straight, because history limits get mistaken for spending limits.

PlanPublished priceExecutionsAI Assistant credits
Self-hosted CommunityFreeNo plan quota; you pay for your own server and opsBring your own key, no credit ceiling
Cloud free trialFree, no card1,000, 5 concurrent, 180-second timeout800, plus 100 AI credits and 20 AI Builder credits
Cloud Starter20€/mo, billed annually2,5002,300/month
Cloud Pro50€/mo, billed annually10,000Up to 13,700/month
Business667€/mo, billed annually40,000; overage 4,000 EUR per extra 300,000, billed 45 days in arrearsNot listed
Startup (self-hosted only)333€/mo, billed annually40,000; eligibility under 20 employees and under €5M fundingNot listed
Cloud EnterpriseContact salesCustomNot included

From the n8n pricing page and the startup plan page, checked September 19, 2026. Three caveats this page will not paper over. Monthly, non-annual prices are not rendered server-side on that page, so no monthly figure is quoted here. The prices served to this check were in euros and may be geo-resolved, so confirm your own currency in checkout. And the overage answer in n8n's pricing FAQ is written for Business and Enterprise — neither that page nor the startup plan page says what happens on Starter or Pro when the execution quota is exhausted.

Meter two: the model bill, and the defaults that multiply it

The formula is simple and the defaults are where the money goes:

cost = production executions × model calls per execution × (input tokens × input rate + output tokens × output rate)

The middle term is the one nobody budgets for, because n8n's billing never shows it. Every figure below was read from the released source at tag n8n@2.39.8, not from documentation.

Node and settingDefaultWhat it does to the bill
AI Agent — Max Iterations10Up to ten model round-trips inside one billed execution; context grows each round, so late iterations cost more than early ones
OpenAI Chat Model — Max Retries2Up to three billed attempts per call; a retried call that produced tokens before failing has usually already cost you
OpenAI Chat Model — Timeout60,000 msA slow call is abandoned at 60s and then retried, so a timing problem presents as a cost problem
OpenAI Chat Model — Maximum Number of Tokens-1 (uncapped)No ceiling on output length unless you set one
Anthropic Chat Model — Max Tokens4,096A hard output cap; cheaper by default, but long generations truncate instead of continuing
Embeddings OpenAI — Batch Size512 documents per requestOne node run can push a very large single request; Kunavo does not serve this node at all (see below)

Sources, all read September 19, 2026: the agent's Tools Agent options, the OpenAI Chat Model node, the Anthropic Chat Model node and the Embeddings OpenAI node. Ten iterations times three attempts is a ceiling, not a typical run — but it is a ceiling that costs thirty model calls and still shows up as one execution on your n8n plan.

A worked n8n AI API cost estimate

These are illustrative token arithmetic, not measured workflow costs and not a bill ceiling. Assume one AI Agent execution that makes three model round-trips — well under the default cap of ten — each sending 6,000 input tokens (system prompt, tool schemas and a scratchpad that grows every round) and receiving 350 output tokens. Rates are live Kunavo catalog prices per million tokens; the quota columns simply multiply by the Starter and Pro execution allowances so the two meters sit side by side.

ModelInput / output per 1MPer execution× 2,500 executions× 10,000 executions
GPT-5.6 Luna$0.07 / $0.42$0.0017$4.25$17.01
Gemini 2.5 Flash$0.09 / $0.75$0.0024$6.02$24.07
Claude Haiku 4.5$0.40 / $2.00$0.0093$23.25$93.00
Gemini 3.8 Flash$0.525 / $2.625$0.0122$30.52$122.06
GPT-5.6 Terra$0.70 / $4.20$0.0170$42.52$170.10
Claude Sonnet 5$2.00 / $10.00$0.0465$116.25$465.00

Two readings. First, the spread between the top and bottom rows under these assumptions is larger than the entire n8n plan fee, which is why arguing about Starter versus Pro before choosing a model is arguing about the smaller number. Second, the iteration ceiling is the real variance: the same Claude Sonnet 5 execution modelled at $0.0465 for three calls becomes $0.4650 if it runs the full ten iterations with every call retried twice — ten times the cost, with no change at all to your n8n execution count. On GPT-5.6 Luna that same three-call execution models at $0.0017.

One rate caveat belongs inside this table rather than under it. Google's own list price for Gemini 3.8 and 3.7 Flash is an introductory rate that runs through December 31, 2026 and doubles on January 1, 2027 (Google pricing, checked September 19, 2026). Kunavo's Gemini 3.8 Flash catalog entry is anchored to that introductory rate and flagged for re-review on the same date, so do not build a 2027 budget on that row without re-checking the catalog. The caveat is specific to those two models; it does not apply to the Gemini 2.5 Flash row.

Scale by your own executions per day before treating any of this as a budget. Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. Cache charges and external tools sit outside this example. The minimum Kunavo top-up is $10 in prepaid credit, which is a funding minimum rather than a subscription or a per-task fee — see billing details.

n8n's "AI credits" are not your model budget

This is the single most-repeated error in published n8n cost writing, and it is worth being blunt about. n8n documents three different credit balances, and the one advertised on the pricing page is not the one that pays for your workflows' model calls.

BalanceWhat it actually pays forWhere it applies
Gateway creditsAI models and tool services used by nodes in your workflowsCloud Starter and Pro only, from n8n 2.36.0
n8n Assistant creditsYour usage of n8n Assistant and the AI Workflow BuilderAll Cloud plans except Enterprise; the 2,300 / up-to-13,700 figures
Free OpenAI API creditsA legacy one-time allowance for new Cloud usersOnly on instances without Gateway credits

The first two columns are quoted from n8n's Gateway credits documentation, checked September 19, 2026; the plan scope for Assistant credits is from the FAQ on the pricing page, which states that AI Assistant is included in all n8n Cloud plans except Enterprise and that unused credits do not roll over. Community threads about "free OpenAI credits in n8n" describe the third, legacy balance and should not be read as current.

Gateway credits are a genuine convenience: n8n routes the node's requests through its own gateway, authenticating with the provider on n8n's behalf, so a workflow runs a model node with no provider account at all. Per n8n's Use Gateway credits page, a node switches between "Use Gateway credits" and "Use my own credential" from one selector, without rebuilding the workflow — though on nodes where only part of the functionality is supported, the unsupported operations are hidden while Gateway credits are selected.

This page cannot give you a numeric Gateway-credits-versus-Kunavo comparison, and neither can anyone outside a Cloud login. n8n publishes its gateway rates only at app.n8n.cloud/service-pricing, behind authentication, and states that rates vary by model and service. There is no published credit-to-dollar conversion. Any article showing you that table built it from something other than a source. What is comparable is structural:

Questionn8n Gateway creditsYour own key (any provider)
Available on self-hosted?No — Cloud Starter and Pro only, and not on Cloud EnterpriseYes, everywhere; no plan gate on the Base URL field
Does the balance expire?Top-up credits expire 12 months after purchaseDepends on your provider; Kunavo credit does not expire on a 12-month clock
Refundable?Top-ups are final; unused credits are not refunded except where law requiresYour provider's terms
If you close the account?Remaining credits are forfeitedYour provider's terms
Who can top up?Instance owner only, and an active paid subscription is required; trials cannot top upWhoever holds the billing account
Scope of the balanceOne balance per Cloud instance, shared by everyone using itPer key or per account, as you structure it
Whose provider account sees the promptsn8n'sYours, or your gateway's

Every row is from the same Gateway credits documentation and its top-up page. The practical consequence: if you self-host n8n, Gateway credits do not exist for you, and bring-your-own-key is not a preference but the only path.

Which route wins, and when

RouteWins whenWhat you give up
n8n Gateway creditsYou are on Cloud Starter or Pro, want zero provider accounts, and value setup speed over price visibilitySelf-hosted and Enterprise cannot use it; a prepaid balance that expires in 12 months, is non-refundable and is forfeited on account closure; rates you cannot read before signing in
Direct vendor APIOne vendor's models run everything, and you want that vendor's own prompt caching and batch discountsA second model family means a second account, a second key and a second balance to watch
An OpenAI-compatible gatewayDifferent workflows want different model families on one key and one balance, and you self-hostVendor-specific extras behind the generic protocol may not be reachable; you set the Base URL by hand once per credential
Local model (Ollama node)High-volume, low-stakes, privacy-sensitive steps where no per-request charge is acceptableCapability gap against hosted frontier models; hardware and operations become your cost

A budgeting note that applies to the three bring-your-own-account rows: n8n gives you no spend cap on model calls billed outside it. The execution quota caps how many times a workflow starts, not what happens inside it. (Gateway credits are the exception — the balance itself stops at zero, and auto top-up takes an optional monthly limit.) If you want a hard ceiling, it has to come from the provider side — a prepaid balance that simply stops when it is empty is the bluntest and most reliable version of that. AI cost optimization covers the rest of the method, and the AI agent API directory shows how n8n sits beside the other clients that take a custom endpoint.

Wiring a third-party endpoint into n8n

n8n ships no generic "OpenAI-compatible" model node. The OpenAI node plus a credential Base URL is the generic route, and the most common tutorial error is looking for that field in the wrong place.

The Base URL lives on the credential, not in the node. In the released source the OpenAI Chat Model node's in-node Base URL option is hidden on node version 1.1 and above, and the Embeddings node's is hidden from 1.2 — so on any node you add today, the in-node override is invisible and the credential is the only place it exists. Neither n8n's OpenAI credential page nor its Anthropic credential page documents the field, but both credentials carry it in source.

n8n credential: OpenAI (Base URL override)
Credential type:  OpenAI  (used by the "OpenAI Chat Model" sub-node)
API Key:          sk-kn-…            # your Kunavo key
Base URL:         https://api.kunavo.com/v1

# Then in the OpenAI Chat Model node:
Model:            ID mode -> claude-sonnet-5
Options > Built-in Tools:   leave OFF on a non-OpenAI endpoint

A source-versus-docs conflict you should know about. n8n's node documentation says the OpenAI Chat Model node defaults to the Chat Completions API. The released source at n8n@2.39.8 says otherwise: responsesApiEnabled has default: true on node version 1.3 and above, so a freshly added node posts to /v1/responses, not /v1/chat/completions. Both are shown here rather than silently picking one; behaviour follows the shipped code, and an endpoint that implements only chat completions will break out of the box until that toggle is turned off. Kunavo's /v1/responses route accepts every enabled chat model, as /v1/chat/completions does, so the node's default endpoint is one that exists here rather than one that 404s — though no workflow has been run end to end to confirm the round trip. See the chat API docs and the OpenAI-compatible API guide.

Three boundaries worth knowing before you build.

BoundaryWhat happensWhat to do
Anthropic credential against Kunavon8n's Anthropic credential authenticates with x-api-key, and both its test and its model dropdown are a GET on /v1/models. Kunavo accepts x-api-key on /v1/messages and /v1/responses but its model list is Bearer-only, so the green-check test fails and the From List dropdown returns nothingUse the OpenAI Chat Model node with the Kunavo Base URL, or set the Anthropic node's model selector to ID mode and type the slug
Google Gemini nodeIt speaks Google's native protocol, which Kunavo does not serveReach Gemini slugs through the OpenAI Chat Model node instead of the Gemini node
Embeddings OpenAI nodeIt honours the credential Base URL, but Kunavo does not serve an embedding model — the route exists and the catalog behind it is emptyKeep a second provider, or a locally run model, for the embedding step of any RAG workflow; see RAG implementation

Also leave the OpenAI node's Built-in Tools off on a custom endpoint. Web Search, File Search and Code Interpreter are that collection in the released source; Conversation ID, Prompt Cache Key and Service Tier sit beside it in the node's Options. All six are OpenAI-platform features; none of them are verified to work through Kunavo, and for models that are translated to chat they cannot. n8n prints its own warning on the node for exactly this situation: when using non-OpenAI models via a Base URL override, not all models might be chat-compatible or support other features such as tool calling or JSON response format.

Set it up and read the first charge

Kunavo has not runtime-tested n8n. There is no published n8n integration page here yet, and every n8n behaviour above is read from the released source at n8n@2.39.8 and from n8n's own documentation — that is a configuration reference, not a compatibility test. Treat the first build as a trial: keep a working route available, run one bounded workflow with a small execution count, then read the charge your account actually recorded for it rather than estimating from the table above.

If you are choosing between gateways rather than between routes, OpenRouter alternatives and the LLM gateway guide cover that comparison, and cheapest Claude API covers the Claude family specifically. When you are ready to fund a key, create a Kunavo account.

FAQ

Does the n8n API cost anything?

It depends which API is meant. n8n's own public REST API — the one that lists and triggers workflows — has no published per-call price; what n8n's documentation states about it is that it is not available during the free trial and that you should upgrade to use it. The AI model APIs your workflows call are a completely separate bill paid to whoever serves the model, priced per token, with no relationship to your n8n plan. When someone asks what n8n API cost is, the answer is almost always the second one.

Do the AI credits on n8n's pricing page pay for my model calls?

No, and this is the most common mistake in n8n cost writing. The 2,300 and up-to-13,700 figures on n8n's pricing page are AI Assistant credits — the allowance for the workflow-building copilot and AI Workflow Builder, refreshed monthly and not rolled over. Model calls made by nodes inside your workflows draw on a different balance called Gateway credits, which n8n documents as a separate product with its own top-up, expiry and availability rules. A third, legacy balance of free OpenAI credits exists only on Cloud instances that do not have Gateway credits. Three balances, three meanings.

What is the best API for n8n?

There is no single winner; the answer follows your deployment and your switching habit. Self-hosted n8n cannot use n8n's Gateway credits at all, so it is bring-your-own-key by definition. On Cloud Starter or Pro, Gateway credits buy convenience — no provider account to open — at the cost of a prepaid balance whose top-up credits expire twelve months after purchase, are not refunded, and are forfeited if you close the account. A direct vendor API wins when one vendor's models run everything and you want that vendor's own caching and batch discounts. A gateway such as Kunavo or OpenRouter wins when different workflows want different model families on one key and one balance. A local model wins for small, private, high-volume steps where no per-request charge is acceptable.

What is the cheapest API for n8n?

Cheapest per token and cheapest to finish the workflow are different questions. Inside an n8n AI Agent node the model decides how many round-trips the run takes: the agent's default Max Iterations is 10, and a weaker model that mis-forms a tool call can spend several of those iterations recovering, so its per-token advantage is multiplied away inside one billed execution. The practical method is to pick the least expensive model that completes your workflow reliably, then read the charge your provider account actually recorded for a bounded test batch. No provider, Kunavo included, can promise the lowest total cost for a task it has not run.

What is the cheapest AI model for n8n?

For deterministic single-shot work — classification, field extraction, routing, short rewrites — the cheapest small model that passes your accuracy check is the right one, because those steps make exactly one model call per execution and the token rate is the whole bill. For agentic steps with tools, judge by finished runs rather than by the rate card, because iteration count is part of the cost. Two structural warnings: an output cap matters as much as the rate, since the n8n Anthropic Chat Model node caps output at 4,096 tokens by default while the OpenAI node ships an unlimited default; and any model priced on a vendor's temporary introductory rate should be re-checked before you build a budget on it.

What is the best model for n8n workflows?

Match the model to the node, not to a leaderboard. A Basic LLM Chain doing extraction is one call and rewards the cheapest model that is accurate enough. An AI Agent with three or four tools rewards reliable tool-calling, because every malformed call costs another iteration of context that grows each round. A summarization step over long documents is dominated by input tokens, so the input rate matters more than the output rate. Note that this question also gets asked in a different sense — which model writes n8n workflow JSON best — which is about authoring, not about what runs inside a live workflow. This page answers the runtime cost question.

Can I point n8n at a custom OpenAI-compatible endpoint?

Yes. In the released n8n source the OpenAI credential carries a Base URL field defaulting to https://api.openai.com/v1, and the Anthropic credential carries one defaulting to https://api.anthropic.com. Neither field is documented on n8n's credential documentation pages, but both exist in source. On nodes added today the in-node Base URL option is hidden, so the override belongs on the credential. One caution: the credential test issues a plain GET on the model list, so an endpoint that does not answer one on the same credential fails the green check even when chat calls would succeed.

n8n plan prices, execution rules, Gateway credits terms and API documentation checked September 19, 2026; node defaults and credential fields read from the released source at tag n8n@2.39.8. Kunavo's Bearer-only model-list behaviour was read from this repository's authentication code, not measured against the live API. No n8n workflow was built or run against Kunavo. Kunavo token rates come from the live catalog, and every dollar figure on this page is illustrative token arithmetic rather than a measured run.