Run a Make AI audit to keep your automations cheaper than your time

If a Make scenario is burning AI credits while you still fix its output, the step is costing more than it saves. Gemini 3.7 Flash's introductory pricing expires on December 31, 2026, and from January 1, 2027, the price will be $1.50 per 1M input tokens and $7.50 per 1M output tokens. A short audit keeps the margin: each step gets the cheapest capable model, the right thinking level, and a cap before the cost becomes routine.

Make made Google's Gemini 3.7 Flash available in Make AI apps in its August 2026 release round. In Make, Gemini 3.7 Flash accepts text, audio, images, code, and video, and returns text. In Make, it provides tool use and control over thinking levels. Google introduced it as a Flash-series model for coding and agents.

The thinking level should match the step

Most scenarios do not need the same model for every step. The highest-burn step is usually the one that runs every day, not the one that looks most impressive. One step may classify an email, another may extract an address, and another may draft a reply. The simplest steps can often run with less reasoning; the harder step may need more. If every step stays on the same setting, you pay for the hardest step in every run.

Look at the input before anything else. A short line from a form does not need the same reasoning as a transcript, screenshot, or file. The output matters too: a short classification costs less than a full draft. A multimodal input can raise the token count even when the answer is short. Check both before changing the model.

The hidden cost is mismatch. A model that overthinks a simple extraction can make a cheap step feel expensive. A model that underthinks a hard step can cost you in rework. The hours you spend watching a scenario fail are the real cost; the credits are just the meter.

The price change is a reminder to audit

The introductory price is $0.75 per 1M input tokens and $3.75 per 1M output tokens. It is half the original Gemini 3.6 Flash cost per million tokens. The deadline is a reminder to look at what each step is actually doing. A price move can turn a step chosen for cheapness into a step chosen for capability, or the reverse. The audit separates them.

The audit should be short enough to repeat. You are checking whether the step still matches the work, not rebuilding the scenario.

The step should earn its model

The test is simple: can the step do the job with less thinking, less output, or a cheaper model? Work through each AI step and answer the same questions.

• List the scenarios: note every Make scenario that uses an AI step.
• Estimate tokens: note input and output length for a typical run.
• Classify the step: simple, complex, or multimodal.
• Choose the model and thinking level: use the cheapest capable option, with low thinking for routine work and higher only when reasoning changes the next action.
• Set a spend cap: stop or reroute when the scenario exceeds expected cost.
• Re-check at price changes: update the model or cap when a model's price moves.

Each item forces a decision. You are asking whether this step needs more than the previous step gave it, not judging the model in the abstract. The checklist keeps the decision local: this step, this job, this cost.

The cap sets a limit. A failing step should stop the scenario and ask for help instead of spending credits on guesses. Keep a note next to each step: what it does, why it uses that model, and what would make you change it. That note is the record. The cap protects the hours you would otherwise spend chasing a failed run.

When a new model appears, do not swap it in everywhere. Swap it into a single step, watch the output, and compare the time you spend fixing mistakes. Keep it when the new model saves attention; leave the old model alone when it only saves a small amount.

Start with the step that runs most often. Keep it when it still needs the bigger model; move it down when it does not, and note the saved attention.