"AI included" on an automation platform is two bills wearing one label. The platform meters its own AI features in credits. The model provider meters tokens. Those are separate ledgers, settled with different vendors, and only one of them is covered by your subscription. Once you can tell them apart, you can work out what an AI workflow costs per run before you build it, and you can usually cut the variable half of that number substantially by changing who buys the tokens.
This matters more now than it did a year ago because the included allowances have been moving. The reference points below were all checked on 5 November 2026.
The two meters
Platform AI credits are the vendor's own currency for the vendor's own AI features. On n8n these are Assistant credits, which fund the in-product assistant that helps you build and debug workflows. On Retool they are pooled AI credits, which cover AI-assisted app building and AI Actions across the whole account. On Make, the base unit is an operation credit consumed by each module action in a scenario, so an AI step costs credits in the same way a Google Sheets row insert costs credits.
Model tokens are what OpenAI, Anthropic, Google or Azure OpenAI charge for the actual inference — input tokens going in, output tokens coming back, plus embeddings if you are doing retrieval. These are billed by the model provider against your own account and key, unless the platform is fronting the model for you and charging credits as a proxy.
The confusion is understandable, because for small volumes of in-editor assistance the platform credit is genuinely all you pay. The moment an AI step runs in production, on every record, the token bill becomes the dominant cost and the credit allowance becomes nearly irrelevant.
Where the allowances currently sit
n8n's published cloud pricing lists Starter at €20 per month billed annually with 2,500 workflow executions and 1,600 Assistant credits, Pro from €50 per month billed annually with 10,000 executions and up to 9,600 Assistant credits depending on plan size, and Business at €667 per month billed annually with 40,000 executions, with the page noting Business is currently self-hosted only. Executions are metered, steps within an execution are not.
Those Assistant credit numbers are lower than they were. From 1 October 2026 n8n reduced included Assistant credits: Starter from 2,300 to 1,600, Pro-1 from 5,700 to 4,000, and Pro-2 from 13,700 to 9,600. If you are reading a comparison article that quotes 2,300 credits on Starter, it predates the change.
Retool's current pricing page lists pooled AI credits alongside workflow runs: Free at 250 AI credits and 500 workflow runs per month, Team at 1,000 credits and 5,000 runs, Business at 3,000 credits and 5,000 runs, and Enterprise at 10,000 credits with custom run limits. Credits are pooled across the account, renew monthly and do not roll over.
Make meters differently again. Each module action in a scenario counts as one credit, so a scenario that reads a record, transforms it, calls a model, updates a CRM and sends an email consumes credits per action rather than per trigger. A fifteen-module scenario costs fifteen times what a one-module scenario costs on the same trigger volume. That makes workflow design a direct line item on the bill in a way it is not on n8n.
The practical consequence: a "10,000 run" plan and a "10,000 operation" plan describe completely different amounts of work. Never compare them as the same unit.
The number that settles the argument
Retool sells additional AI credits in packs of 1,000 credits at US$100 per pack. That works out to ten US cents per credit, which is the clearest available statement of what platform-metered AI is worth to the vendor.
Set that against what the same work costs when you buy the tokens yourself. For most production automation steps — classify this email, extract six fields from this PDF, draft a reply for human review — the raw model call on a current mid-tier model costs a small fraction of a cent to a few cents, depending on prompt length and model choice. The platform credit is a convenience wrapper with a margin on it. For building apps and workflows, that convenience is worth paying for. For a step that runs ten thousand times a month against customer records, it generally is not.
Also worth noting that the model side of the market is getting more granular, not less. Make's 2026 documentation records that from 25 August 2026 input and output model pricing are separated rather than blended, which lowers the cost of workflows that send long prompts and receive short answers — which is most classification and extraction work. Any cost model built on a single blended "per AI call" figure is now understating the difference between workload shapes.
Bring your own key as the production default
The rule that holds up across all three platforms: use platform AI credits for building, use your own provider key for anything that runs in production.
Build-time assistance is bursty, low-volume and genuinely useful. Production inference is high-volume, repetitive and the thing that will surprise you on an invoice. Separating them gives you four things that the pooled-credit model does not.
Price visibility. You see input tokens, output tokens and model name in the provider's own usage console, broken down by API key. You can attribute spend to a specific workflow if you issue a key per workflow.
Model choice. You can run a cheap model for triage and an expensive one only for the cases that fail triage. Pooled credits usually flatten that distinction or tie you to the platform's model selection.
Spend caps. Provider consoles let you set hard monthly limits and alerts. A runaway retry loop hits a ceiling instead of quietly consuming a credit pool and then silently failing mid-month when it empties.
Data handling answers in writing. When the model account is yours, you control the retention and training settings, and you can point at a contract when someone asks whether customer data is used for training. Choosing the right provider, region and key arrangement — and wiring the keys, caps and fallbacks into the workflow itself — is the substance of AI integration work rather than an afterthought once the automation is built.
What bring-your-own-key commits you to
It is not free in effort, and anyone telling you otherwise has not run one in production.
- Secrets management. Keys live in the platform's credential store, get rotated on a schedule, and are revoked when staff leave. Someone owns that calendar.
- Rate limits and retries. Provider rate limits are per-account. Two busy workflows sharing one key will throttle each other. Retries on a 429 cost tokens every time they run, so a retry policy without a backoff and a ceiling is a cost bug.
- Model version drift. Providers deprecate model versions. Pin the version in the node, and own the task of testing and moving when the pin expires.
- Failure behaviour. Decide what the workflow does when the model is unavailable: queue, fall back to a smaller model, or route to a human. Silent skips are the expensive option because nobody notices for a fortnight.
- Logging. Log prompt length, response length, model and latency per run. Without it you cannot explain a bill increase, and bills increase for reasons unrelated to volume — a longer prompt template, a chattier system message, a schema change that doubles the input.
- Residency. Self-hosting the automation platform does not by itself keep data in Australia. Retool's pricing material indicates cloud, self-hosted and air-gapped deployment options on higher plans, which establishes deployment choice but says nothing about where the model provider processes the prompt. If the data cannot leave the country, the model endpoint has to be the thing you check, not the orchestrator.
Reading an Australian price off a European page
n8n's plan prices are published in euros and Retool's credit packs in US dollars. The pricing pages reviewed do not establish an Australian landed price, GST treatment or billing entity for any of the three platforms. Before committing to an annual plan, confirm through the Australian checkout or a sales quote: the display currency, whether GST is included or added, the card issuer's conversion method, the contracting entity, and the renewal and cancellation terms. An annual commitment priced in euros is a foreign-exchange position as well as a software subscription.
Likewise, do not treat self-hosted n8n as free. The Community Edition can be self-hosted indefinitely, but features such as SSO, environments and projects sit behind paid plans, and the server, backups, upgrades, monitoring and security review are now your line items rather than the vendor's.
Build the cost model one workflow at a time
Pick the single AI workflow you expect to run most often. Write down four numbers: how many times it runs per month, how many model calls happen inside one run, the approximate input and output length of each call, and which platform unit each step consumes. That gives you a monthly token estimate you can price against your provider's published rates, and a separate count of platform executions, operations or runs you can check against your plan allowance. Most of the time one of those two lines is an order of magnitude larger than the other, and that tells you where to spend your attention.
If you would rather have the arithmetic done properly across the whole portfolio — platform subscription, token spend, hosting, build and ongoing maintenance, against fully loaded labour cost — run your numbers through our automation ROI calculator and send us the result. We will come back with a scoped build estimate in the two to four week range we normally work to, a monthly running cost split between platform and model spend, and a straight answer on which of your candidate workflows pay for themselves and which ones do not.