Should you build your coaching layer on a frontier model yourself?
If LLM infrastructure is your core competency, build. You will get exactly the behavior you design, you keep every prompt, and nothing on this page argues otherwise.
For everyone else, the trap is that the model call is the easy 10%. The other 90% is output contracts, deterministic validators, a scoring harness, and standing measurement across model migrations. We did not estimate that 90%: we measured it on production traffic, and the numbers below come from that work.
Build it yourself when, buy the layer when.
| Build on the model directly if | Use a coaching layer if |
|---|---|
| You have a team whose job is prompts, evals and model operations. | Your team ships app features, and coaching is one of them. |
| Coaching logic IS your product’s moat and you must own every line of it. | Coaching is table stakes your users expect, not your differentiation. |
| Your scale makes per-call economics dominate engineering cost. | You are at the scale where four months of engineering costs more than years of API fees. |
| Your data may not transit any third party, contractually. | A stateless per-call proxy that stores nothing satisfies your data posture. |
What you actually rebuild, with the measured evidence.
Each line below exists because production traffic proved it necessary. The full detail, including where our own numbers are weak, is on the benchmark page.
Output contracts per endpoint
The import contract alone moved adversarial-document fidelity from 78.2% to 96.5% on the same model; structured intensity (tempo, RPE, percent 1RM) went from 13% preserved to 100%.
Deterministic validators behind the model
A coherence guard now blocks a next step that raises intensity after an injury warning (found twice in 183 real analyses). A length cap absorbs raw-model overshoot. Zero banned cliches in 284 production analyses is a filter, not luck.
A scoring harness you can trust
Four bugs in our own instrument produced false failure rates up to 67% before correction. Building the benchmark was as hard as building the prompts.
Standing measurement across model migrations
One migration silently fixed a 65% unsolicited-action defect; another silently multiplied latency by 5. Neither is visible without a suite that replays real traffic.
Domain post-processing
PR detection and session classification are computed, not asked of the model. The model does interpretation; arithmetic stays deterministic.
Raw inference is not where the money goes.
Measured per-call token averages from our scored runs, at public list rates of $2 in / $12 out per million tokens:
So a build decision priced on tokens is priced wrong in both directions: tokens are cheaper than most teams expect, and the engineering loop around them is more expensive. The full per-unit math, including our own prices next to every alternative, is in how much a fitness coaching API costs.
Pushback we agree with.
Frontier models keep improving. Does the layer become worthless?
Model upgrades genuinely absorbed one of our defects: a migration took unsolicited actions from 65% to zero, and we published that. What upgrades do not absorb: contracts (a better model still needs to know your schema), safety coherence rules, and the measurement that tells you what a migration changed. Better models make the layer thinner, and they make knowing-what-changed more valuable, not less.
Is a wrapper around someone else’s model defensible at all?
A naked wrapper is not, and the market is right to be cynical about them. What we sell is not model access: it is eight endpoint contracts extracted from a shipped training product, the validators behind them, and a production benchmark with our weaknesses printed on it. Judge whether that clears the bar on the numbers.
We already prototyped coaching with a system prompt. It looks fine.
Ours looked fine too. Then reading 183 real outputs found two analyses that told an athlete they were at elevated injury risk and, in the next sentence, to push intensity. One percent of outputs, invisible in any demo, and exactly the kind of thing that ends up in a screenshot. Fine-in-a-demo and safe-at-volume are different claims; only measurement separates them.
What does it cost to build a coaching feature in-house?
Mostly engineering time, not model calls. A realistic first pass at output contracts, validators and an eval harness for a handful of endpoints takes about 4 months. Even priced at one engineer on the US median software developer wage ($135,980 a year in May 2025, per the Bureau of Labor Statistics), that is about $45,000 in salary, or about $65,000 once benefits are counted, because wages are 70% of what a US private employer pays per hour. The same money buys more than 2.3 million raw model calls at our dearest measured per-call cost. And the build is never finished: every model migration needs the measurement run again.
If we buy now, are we locked in?
Less than with most vendors, by design. Each call carries the athlete context it needs in the request body, so your users, history and programs stay in your database, and there is no sync to unwind. Endpoints are versioned under /v1/, and breaking changes ship as new versions with a 12-month deprecation window. Plans are monthly and cancel anytime. If you later decide to build, keep our response schemas as the contract your own implementation has to meet.
Does the EU Artificial Intelligence Act apply to a coaching feature in a fitness app?
Yes, but only its light transparency rule, and the same rule applies whether you build or buy. Fitness and training coaching is not one of the high-risk uses listed in Annex III of Regulation (EU) 2024/1689, so the heavy conformity obligations do not apply, unless your app is itself regulated as a medical device, which the Act covers separately. What does apply, since 2 August 2026, is Article 50: a system that interacts directly with people must be designed so they are told they are dealing with an artificial intelligence system, unless that is obvious from the context. Breaches can be fined up to 15 million euros or 3% of worldwide annual turnover. In practice, say on the screen where the athlete reads the coaching that it is generated by software, not written by a human coach. That label belongs in your app either way, because your app is where the athlete meets the coaching.
Does a coaching feature in a US fitness app need FDA clearance?
Not while it stays a fitness feature, and that holds whether you build or buy. Since the 21st Century Cures Act of 2016, section 520(o)(1)(B) of the Federal Food, Drug, and Cosmetic Act has excluded from the definition of a medical device software intended for maintaining or encouraging a healthy lifestyle that is unrelated to the diagnosis, cure, mitigation, prevention or treatment of a disease or condition. The FDA guidance General Wellness: Policy for Low Risk Devices, revised on 6 January 2026, names physical fitness, weight management and sleep management as general wellness uses, with claims to log or track exercise, or to improve endurance or strength, as examples. It also says a non-invasive wearable estimate such as heart rate variability can count as a wellness output when it is intended solely for wellness use and meets its other conditions. What changes the answer is the claim, not the technology: a claim to treat or diagnose obesity, or to treat an eating disorder, is not a general wellness claim. In practice, keep coaching copy about training, recovery and habits, never about diagnosing or treating a condition, and point users to a professional for health concerns, which our terms already require of every app that uses CoachLayer.
Wage figures: BLS, Software Developers (May 2025 median) and BLS, Employer Costs for Employee Compensation (private industry, June 2026), checked October 2026. EU rules: Article 50 and Annex III of Regulation (EU) 2024/1689, checked November 2026. US rules: the FDA guidance General Wellness: Policy for Low Risk Devices (6 January 2026), checked November 2026.
Skip the 90%, keep the proof.
Eight production-tested endpoints behind one key, with the benchmark to hold us to it.
Get a sandbox key