Our verdict: start with a cheap model and one tool
An agent costs more than a fixed scenario because it spends tokens on reasoning and may call several tools per request. [3] The cost is controllable: model choice changes token credits by more than ten times, and Make's own guidance is to test whether a cheaper model meets your goal before moving to a more expensive one. [4] [5]
Who should use an agent: a workflow that receives unstructured input, such as free-text emails, and needs judgment to pick the next step. Who should not: a workflow whose steps are always the same; a normal scenario is cheaper and easier to audit. Make itself says agents work alongside deterministic logic, not instead of it. [2]
What an agent is made of
Make's documentation lists five parts. The LLM is the reasoning engine. Instructions are the system prompt that sets role, limits and process. Context is files in long-term memory, such as policies. Tools are the modules, scenarios and MCP tools the agent may call. Reasoning is visible when a reasoning model decides deeper thinking is needed. [1]
Make's product page says agents are built with Make's visual builder, act as reusable building blocks, and show their decisions in a Reasoning panel. [2] The introduction page says the app is in open beta and that both functionality and pricing may change; that page is marked as covering a previous version, so check the current AI Agent (New) documentation too. [1]
How agents use credits
From Make's credit usage page for the AI Agent app, read on 2026-10-29.
| Action | With Make's AI Provider | With a custom AI provider connection (paid plans) |
|---|---|---|
| Run an agent | 1 credit per operation plus token credits [3] | 1 credit per operation; tokens billed by your provider [3] |
| Chat | 1 credit per operation, plus 1 per operation from called tools, plus token credits [3] | 1 credit per operation plus 1 per operation from called tools [3] |
| Tool call | 1 credit per operation plus token credits [3] | 1 credit per operation [3] |
| Knowledge upload, PDF or DOCX | 1 credit plus 10 tokens per page, plus description and embedding tokens [3] | 1 credit plus embedding tokens [3] |
| Query knowledge | 1 credit plus embedding-token credits [3] | Same [3] |
What a credit buys by model
Make publishes how many tokens one credit covers on its AI Provider, and notes the rates may change as models change. The rows below are a selection from the full list on the credits page, read on 2026-10-29. [4]
| Model option | Input tokens per credit | Output tokens per credit |
|---|---|---|
| Small and Medium (GPT-5 nano) | 18,080 | 2,260 |
| Large (GPT-5 mini) | 3,616 | 452 |
| Claude Haiku 4.5 | 904 | 181 |
| Claude Sonnet 4.5 | 301 | 60 |
| Claude Opus 4.8 | 180 | 36 |
Worked example: 200 agent runs a month
Hypothetical: an agent reads a support email (3,000 input tokens including instructions), writes a 500-token reply draft and calls two tools: look up the order, then create a draft. We assume input and output token credits are added together and ignore the tools' own token use.
On Large: 3,000 / 3,616 + 500 / 452 is about 1.9 token credits, plus 1 credit for the run and 2 for the tool calls, so about 4.9 credits per email, or about 990 credits for 200 emails. [4] [3] On Claude Sonnet 4.5: 3,000 / 301 + 500 / 60 is about 18.3 token credits, so about 21.3 per email, or about 4,260 for 200 emails. [4]
The first fits inside the Free plan's 1,000 credits with almost no headroom; the second needs a paid tier. With a custom provider connection on a paid plan, Make would charge about 3 credits per email (run plus two tools) and your provider would bill the tokens. [3]
Checklist for a safe first agent
- Start with one tool and a narrow job; add tools only after the first works.
- Give the agent the minimum data it needs, and assume anyone could see what you share with it. [5]
- Do not rely on prompt rules to protect sensitive data. [5]
- Write tool names and descriptions that say when to run them, including prerequisites such as a required ID. [5]
- Test with the cheapest model first and compare outputs before upgrading. [5]
- Leave the Thread ID blank unless you need conversation history, and cap output tokens. [5]
- Add a manual approval before any action that sends, pays or deletes. [2]
- Review the token usage summary for the first 20 runs and project the monthly credits. [5]
Watch demos / Official walkthroughs
See how the tools work.
Videos load from YouTube only when you press play. These are provider videos, not our own product tests.
Make / Official video
Make—Building Your First Automated Workflow from Scratch
What to look for: Follow how modules connect and how data moves through the first workflow.
Watch on YouTubeQuestions before you decide
Are Make AI Agents available on the Free plan?
Make says the app is available on all plans using Make's AI Provider; custom AI provider connections require a paid plan. [1]
Do I pay twice if I use my own OpenAI or Anthropic account?
You pay Make 1 credit per operation and pay your AI provider directly for tokens, instead of paying Make token credits. [3] [4]
How do I limit what an agent can do?
Make suggests giving the agent minimal access to sensitive data and not relying on prompt rules alone, and its product page says you can add manual approvals or stop the agent at specific points. [5] [2]
Sources & verification
Product details and prices can change. Check the linked provider before buying.
- Make Help Center: Introduction to AI agents Accessed 29 Oct 2026
- Make AI Agents product page Accessed 29 Oct 2026
- Make Help Center: Credit usage for AI agents Accessed 29 Oct 2026
- Make Help Center: Credits Accessed 29 Oct 2026
- Make Help Center: AI agent best practices Accessed 29 Oct 2026
Sources link directly to providers. Product buttons may use separately labeled affiliate links. Read our disclosure.