Make AI Agents and the Make AI provider: how token-based credits work, and when to use your own key

Make now lets you add AI to a scenario in two ways: through its own built-in AI provider, which needs no outside account, or through your own OpenAI, Anthropic or Gemini account. Both work. They are billed very differently, and the difference shows up in your credit usage a few weeks after someone adds "summarize this email" to a busy scenario.

The direct answer: with Make's AI provider, Make converts the tokens each AI step uses into credits, at a rate that depends on the model, so a large model can use dozens of times more credits than a small one for the same prompt. With your own key (paid Make plans only), Make charges its normal one credit per operation and your AI provider bills you separately for tokens. Make's provider is simpler and fine for light, small-model use; your own key usually makes more sense for heavy use, larger models, or when you need your own contract with the AI vendor.

The two ways to run AI in Make

Make's credits help page sets out the options, as published in October 2026:

• Make's AI provider (all plans). Used by Make AI Agents and the Make AI Toolkit. Make connects to the model for you, so you do not need an OpenAI or Anthropic account. You pay Make in credits, based on tokens and operations.
• Custom AI provider connection (paid plans only). You link your own OpenAI, Anthropic or Gemini account with an API key. Make charges credits for operations, and you pay your provider directly for tokens.
• Automatic connections. A few modules (Make AI Web Search, Make AI Content Extractor and the Simple Text Prompt modules in the OpenAI, Gemini, Groq and Claude apps) use a provider Make selects, and their credits include token usage. The same page warns that Content Extractor modules may use more than one credit per operation.

The newer Make AI Agent app is labeled open beta in Make's AI Agent (New) documentation, which says product functionality and pricing may change. It follows the same split: Make's provider on all plans, other providers on paid plans, plus an optional fallback connection if the primary one fails.

How token-based credits are calculated

A token is a chunk of text; Make's page says one token is about four characters of English, and around 100 tokens make a paragraph. Every AI step uses input tokens (your instructions, the email or document you pass in, any conversation history) and output tokens (the answer).

With Make's provider, each model has a conversion rate: how many input tokens and how many output tokens equal one credit. A selection from the credits help page, as listed in October 2026:

Make says these models and rates may change as new models are released, so check the page before you estimate.

A worked example

Take one AI step that reads an incoming client email of about 2,000 input tokens (instructions plus the email) and writes a 300-token summary. Using the rates above, the token part alone comes to roughly:

• Small model: 2,000 ÷ 18,080 + 300 ÷ 2,260 = about 0.24 credits.
• Claude Sonnet 5: 2,000 ÷ 452 + 300 ÷ 90 = about 7.7 credits.
• GPT-5.5: 2,000 ÷ 180 + 300 ÷ 30 = about 21 credits.

Run that 1,500 times a month and the gap is several hundred credits against tens of thousands, for the same scenario. Make's page bills on tokens and operations, so treat these as estimates and check the credits shown on a few real runs before you switch the scenario on. Our guide to Make.com credits covers the non-AI side of the bill.

Why AI agents use more than you expect

An AI agent is not one call. Make's AI Agent documentation describes a setting for the maximum steps per agent call: the number of times the agent may call the model to answer one request. Each step re-sends context, so usage grows with:

• The number of tools the agent can choose from, because their descriptions are part of the input.
• Knowledge files and long inputs, such as attached PDFs.
• Conversation history, if the agent keeps it.
• Reasoning effort, which the documentation says uses more AI tokens when set higher.

The same documentation mentions that prompt caching can reduce credit usage, and that a step timeout defaults to 5 minutes, with a maximum of 10. Set the step limit and history length deliberately rather than leaving them open.

When Make's provider makes sense

• You are on Make's free plan, where it is the only option.
• AI use is light: a few hundred short classification or extraction runs a month on a small model.
• You do not want another vendor account, invoice and API key to manage.
• You are testing an idea and want one bill while you find out whether it is useful.

When to connect your own key

• Heavy or growing volume. Your credits then pay only for operations, and you see token costs on your AI provider's own usage dashboard.
• You need a large model. At the higher-end conversion rates, a single step can use more credits than the rest of the scenario.
• You already pay for API access with OpenAI, Anthropic or Google for other work.
• You need your own terms with the AI vendor, for example around data retention, a data processing agreement or a specific region. With your own account, the contract for the model call is between your firm and that vendor; check with your lawyer or compliance lead what it covers before client data goes through it.
• You want a second brake on spending. Check your AI provider's console for usage or spending limits; where they exist, they work alongside Make's own credit alerts.

The trade-off is administration: a key to store safely, a second bill, and a connection that must belong to the firm rather than one employee. Our article on automations when an employee leaves explains why that ownership matters.

Practical rules for AI steps in Make

1. Start with the smallest model that does the job, and only move up when results are not good enough.
2. Trim the input. Pass the email body, not the whole thread with signatures; pass the fields you need, not the full record.
3. Filter before the AI step, so only items that need AI reach it.
4. Cap agent steps and history, and give agents only the tools they need.
5. Turn on credit usage alerts in Make and spending limits with your AI provider.
6. Review monthly. Compare credits used by AI modules with the value they deliver, as you would any other tool. Our guide on where AI assistants fit in a small firm helps decide which tasks deserve AI at all.

How Flow Mesh helps

Our stack audit starts after a first call and is scoped in writing before it begins. It reviews the automation and AI tools you already pay for, including which scenarios use the most credits and whether a different model or connection would serve better, and gives each tool a verdict: keep, configure, consolidate or replace. Any rebuild is quoted in writing after the audit and carried out inside your own accounts.

Key takeaways

• Make's AI provider works on every plan and bills tokens as credits at a per-model rate; your own key (paid plans) costs one credit per operation plus your provider's token bill.
• Model choice dominates the cost: the same step can use well under one credit on a small model and over 20 on a large one.
• AI agents call the model several times per request, so tools, knowledge files, history and reasoning effort all add tokens.
• Use Make's provider for light, small-model work; connect your own key for heavy volume, large models or when you need your own vendor terms.
• Start small, trim inputs, filter first, cap agent steps, and set alerts on both sides.