Compare Make runs using workload-adjusted measures, not raw credit totals alone. Record inputs, accepted outputs and feature mix first, then calculate usage per useful result for genuinely comparable runs.
Understand why totals can mislead
Make documents usage views at several levels, including per-run, scenario and builder/module. Its credits guide says non-AI apps generally use one credit per operation, while AI features can vary with connection type and token or other usage factors.
Consequently, equal operation counts do not establish equal credit consumption across every feature type. Nor does a smaller total establish greater efficiency if a run completed fewer useful outputs.
Keep provider-managed and custom-provider AI connection paths distinguishable in your review notes where relevant.
Build a normalised run sheet
For each run, record:
- Review window and process version.
- Input count and input complexity band.
- Accepted output count and rejected output count.
- Observed credit usage.
- Non-AI and AI feature mix, including connection path.
- Exceptions, repeat processing and reviewer minutes.
Define “accepted” before comparison. For a hypothetical enquiry-classification process, it might mean an output with every required field and a reviewed destination. It should not mean merely that a response exists.
Use synthetic records when constructing a controlled comparison; do not copy customer material into a benchmark.
Calculate two useful denominators
Hypothetical run A consumes 300 credits for 100 inputs and 90 accepted outputs. Run B consumes 240 credits for 60 inputs and 60 accepted outputs.
A uses three credits per input and approximately 3.33 per accepted output. B uses four credits per input and four per accepted output. B has the lower total but the higher observed usage per accepted result.
That still does not prove A is preferable. Its acceptance rate is 90%, versus 100% for B. If rejected outputs require substantial manual repair, A may create more total work.
These figures are invented observations for arithmetic, not documented Make consumption rates.
Separate feature mix before judging
Divide comparisons into similar workload groups: short versus long inputs, ordinary versus exception cases, and different AI connection paths where applicable. Do not average away a material difference in the work being attempted.
If one run performs an additional review stage, note that added outcome explicitly. Removing the stage merely to lower usage may reduce quality rather than improve efficiency.
Use the module-level evidence to investigate unexplained differences after normalisation, not to manufacture a universal cost-per-operation assumption.
Decide whether the comparison is valid
Stop the ranking if acceptance criteria, process versions or feature mix differ materially. Next, run a controlled synthetic batch under the same requirements and compare both usage and reviewer effort.
For access options, check the automation catalog, reviewing access format and saved terms before buying. Marketplace account/code/token offers are not proof of official subscription equivalence.
Sources and scope
Source material checked 29 October 2026; features/policies may change. Make usage tracking establishes review levels; Make credits establishes feature-dependent credit behaviour. The worksheet, denominators and examples are editorial advice.