What do our AI agent trials cost? A campaign summary

One internal campaign covered eight trials and the first set of features for each project. Here is what its $105.64 includes, and why that is not the price of an application.

On August 10, 2026, our campaign record totalled $105.64 in AI usage across eight trials. Each trial started with an idea and built an initial set of features. These were experimental development costs, not eight complete applications delivered to customers.

What we asked the agents to build

The subjects covered inventory, a library, file organisation, shared expenses, room bookings, working hours, habits and invoices. The process included defining the need, describing the rules, building the first set of features and checking it, with repairs when triggered.

An “initial set of features” means that some planned functions remain to be built. For example, the room booking trial did not deliver cancellations. Reaching the end of the trial process therefore does not mean the application is ready for everyday use.

What the $105.64 measures

The total adds together the costs recorded during those eight executions. None was stopped by its spending limit. The figures cover AI calls recorded by our test setup; this page summarises them rather than providing supplier invoices or a complete breakdown.

The judges review the results after execution and are not included in this total. An interrupted first attempt, recorded separately in the campaign report, is also excluded. The figure therefore does not represent the full research cost.

What is outside that amount

Our time preparing, analysing and correcting the engine is not valued in this record. It also excludes launching a customer product, hosting it, external services, maintenance and user support.

To compare it with a quote, you need to include the same work over the same period. The application budget guide gives a fictional example separating construction, launch and operating expenses.

Low spending can hide an incomplete result

We also found faults in inexpensive trials: a check reported a satisfactory result without properly replaying the application’s startup. Spending less is useful only if the expected result works.

Our account of the comparison campaigns explains how the checks evolved. The detailed reports remain internal; the figures published here are our summary and do not constitute an independent audit.

Budgeting your own trial

Describe one feature to try, set a budget and record spending alongside the result obtained. Keep AI costs, outside help and your own time separate. Then add the expenses needed to use the product over time.

Maestro is free during its invitation-only Mac beta; AI usage is paid for separately. The five-year calculation helps you think about maintenance and subscriptions beyond the initial build.