How iAM360 made its AI coach dramatically more efficient without lowering the quality bar

iAM360 is a fitness and wellness platform that helps people understand their wearable, workout, nutrition, sleep, and recovery data. Its AI coach turns that information into practical answers: how hard to train, when to recover, what patterns are affecting progress, and what to do next. A connected trainer platform helps coaches manage clients and create programs.

That means the AI handles very different kinds of work. Some requests require serious reasoning, such as analyzing weeks of health data or building a personalized plan. Others are much simpler: identifying the date of a meal photo, classifying a message, or returning information in a predefined format.

Originally, the system was modeled as if every task used gpt-5.6-terra with its highest reasoning setting, xhigh. For the hardest analysis, that pairing is exactly right. As the default for every request, it is expensive precision applied to work that never asked for it.

Using rightmodeler’s routing and evidence framework, iAM360 taught the system to assign each job to the appropriate level of AI.

The counterintuitive part: right-sizing went in both directions. Some complex coaching paths were upgraded from Terra to flagship Sol at the same time narrowly defined routine tasks moved to smaller, faster models. The routing changed more than the models, too:

The comparison baseline is a hypothetical starting point where every request uses Terra with xhigh reasoning. At identical text usage, smarter routing alone reduces the modeled AI cost by approximately 25 to 26%. The larger saving comes from avoiding excessive reasoning on simple tasks.

In a representative request, the routed system costs approximately 3.1 cents where the all-Terra xhigh version costs approximately 7.25 cents: a modeled saving of roughly 56 to 57% per request.

The honest caveat is that the headline depends on how much extra output and reasoning xhigh actually produces. Even in the most conservative case, identical output volume, routing alone still saves 25 to 26%:

The smaller models are not trusted with every request. They are assigned narrowly defined tasks with predictable outputs and automated validation. The most difficult work still receives a flagship model, and some complex paths were upgraded from Terra to Sol.

At the same representative token shape, the savings compound:

This is what rightmodeler produces: a routing policy where every job runs on the smallest model and reasoning effort that holds its quality bar, with the flagship kept, and sometimes promoted, for the work that deserves it.

Brian Douglas, Founder, iAM360