Make's AI Steps Now Cost Less to Read, More to Write
Make now bills its built-in AI separately for input and output tokens. Steps that mostly read got far cheaper; steps that mostly write got dearer — on the Large tier by up to 3.3 times, the ceiling a step approaches as its output dwarfs its input. Free plans can't route around it.
Original announcement
Make — Aug 25, 2026What happened
On 25 August, Make changed how its built-in AI Provider turns tokens into credits. It used to charge by total tokens, input and output combined. It now charges input and output at separate rates — "input tokens are significantly cheaper and output tokens are moderately more expensive," in Make's words.
The rates, from Make's own release note, effective 25 August 2026:
| Model tier | Before: total tokens per credit | Now: input tokens per credit | Now: output tokens per credit |
|---|---|---|---|
| Small | 5,000 | 18,080 | 2,260 |
| Medium | 3,500 | 18,080 | 2,260 |
| Large | 1,500 | 3,616 | 452 |
This only affects Make's own AI Provider. If you connect your own OpenAI or Anthropic key, nothing changed — you pay that provider directly.
What a step pays now depends on its shape
Reading got cheaper and writing got dearer, so no single percentage describes this change. Every AI step lands somewhere between two limits, and where it lands is set by how much it writes compared with how much it reads.
We worked the limits out from the rates above, effective 25 August 2026:
| Tier | A step that only reads | A step that only writes | Gets cheaper while its output stays under |
|---|---|---|---|
| Small | −72% | +121% (2.2×) | 0.6× its input |
| Medium | −81% | +55% (1.5×) | about 1.5× its input |
| Large | −59% | +232% (3.3×) | a quarter of its input |
Read those as limits, not quotes. No real step is pure reading or pure writing, so every step falls between the two columns. The more a step writes relative to what it reads, the closer it moves to the right-hand column — and no step can go past it.
Here's where five ordinary steps land. The token counts are illustrative; the rates are Make's.
| Step | Tokens in / out | Medium | Large |
|---|---|---|---|
| Extract fields from a document (Make's own example) | 2,000 / 100 | 0.60 → 0.15 credits (−74%) | 1.40 → 0.77 (−45%) |
| Classify an email | 800 / 20 | −77% | −51% |
| Summarise a thread | 3,000 / 300 | −68% | −32% |
| Draft a reply | 500 / 400 | −20% | +71% |
| Write a draft from a one-line prompt | 50 / 1,500 | +50% | +222% |
(Make's release note gives its extraction example as a 73% reduction; the arithmetic comes to 74%, and the difference is rounding.)
So an automation that pulls invoice numbers out of emails just got much cheaper. One that turns a one-line brief into a 1,500-token draft on the Large model now costs about 3.2 times what it did — close to that tier's 3.3× ceiling, because almost everything it does is writing.
Small and Medium now cost exactly the same per token. Before, Small was the budget option. There is no longer a price reason to pick it over Medium.
If you're on Make's free plan
Make's credits documentation says its AI Provider "is available to users on all plans", but "custom AI provider connections are available to users on paid plans." On Free, the built-in provider is the only way to run an AI step — so if your free-plan automation drafts text, you absorb the increase with no way around it inside Make.
What to do
Open any scenario with an AI step and check two things: which model tier it uses, and whether it mostly reads or mostly writes.
- Reading steps — extract, classify, summarise, route: leave them. They got cheaper on every tier.
- Writing steps on Large: test whether Medium is good enough. On a short-prompt, long-output step the gap is now wide.
- Writing steps at volume on a paid plan: compare against connecting your own API key, which bypasses Make's credit rates entirely.
We've written before that Make's native AI modules bill by token rather than per run, which is why a single AI step can cost many credits. That's still true — what changed is which half of the tokens you're paying for. And if those steps pass client data to a model, what your plan lets the model provider do with it is a separate question worth answering first.