What our ×10 measures: instruction text, not your bill

Our historical measurement compares the instruction volume of two methods. It does not mean ten times cheaper or ten times better. Here is the calculation and its scope.

Before building a feature, an AI agent receives instructions: its role, the expected result and the checks to perform. We counted this text to understand the space taken up by the method itself. The ×10 figure comes from that count, made on the versions studied in July 2026.

What goes into the calculation

The compared cycle runs from an idea to a first verified development task: framing, specification, architecture, task breakdown, development and checking. For Maestro, we reconstruct the instructions sent by the engine. For BMAD-METHOD v6.10, we count the files that the corresponding steps must load.

Project code, AI responses, user messages and documents produced during the work are excluded on both sides. We therefore count only part of what the agents actually read. Optional BMAD files are also excluded.

The result in its original unit

We also estimated these volumes at around 6,400 and 65,700 tokens by dividing the character count by four. A token is a piece of text processed by the AI. This conversion remains approximate, particularly because the instructions being compared are in French and English.

A cumulative total, not memory at every moment

The two methods do not divide up the work in exactly the same way. We matched steps that serve comparable functions. BMAD loads some files progressively, so the cumulative total does not describe the amount of text present at a particular moment.

The calculation shows a difference in instruction volume over that historical workflow. It does not prove that a shorter method understands your request better, works faster or produces better software. Nor does it describe every version released since that measurement.

Why the bill is a different question

Billed usage also depends on responses, files read, retries and the model used. The ×10 ratio cannot be used to divide a project budget by ten. Our article on AI spending presents usage observed in other tests, with its own limitations.

Quality scores follow a separate protocol again. The campaign on 6 August 2026 produced 89.5 versus 83.5 across nine scenarios, limited to a first development batch. Those scores do not follow from the instruction count: the campaign review explains how they were obtained.

How to read a numerical claim

Always ask what is counted, in which unit, on which versions and with which exclusions. When choosing a tool, also try it on your real needs: a character count cannot replace a working feature. Our account of the method change explains the decisions these tests informed.