Fawad is CEO and co-founder of Penguin Ai, a healthcare AI startup that sells to payers, providers and revenue cycle management companies.
Ask a health system which AI model reviews its prior authorizations and codes its claims, and it’s usually one of the biggest models on the market, the one built to write code and pass the bar exam. It’s also the tool deciding whether a knee replacement gets approved.
Nobody sat down and chose to run this kind of routine, high-volume work on the most expensive AI available. It happened because that model was already integrated, and it worked. No one revisited the cost of keeping it running. A year later, health systems are paying frontier prices for what is mostly piecework.
In July, more than 75 AI and technology companies, including OpenAI, Google, Microsoft and Nvidia, signed a letter arguing that America’s AI advantage depends on giving every organization access to open-weight models suited to the job at hand. When the labs building frontier models say that, healthcare should ask why it’s still running everything through the most expensive model on the market.
Small Models Are Winning
Physicians complete an average of 40 prior authorizations per week, consuming 13 hours of physician and staff time. Multiply that across coding queues, claims adjudication, denial follow-up and the daily pile of faxes and scanned notes still waiting to become structured data, and you get high-volume, repetitive work.
A study that fine-tuned Llama models as small as 3 billion parameters to sort clinical notes into their sections found F1 scores between 0.90 and 0.94, well ahead of GPT-4o’s 0.72. A separate finding makes the same point through cost. An open-weight model, Mistral-Large, matched GPT-4o almost exactly on pulling findings out of radiology reports, 0.926 to 0.924. Both scores were far ahead of an older rule-based system’s 73.1, and the open-weight model got there running at a fraction of the size.
Amazon’s own CTO, Werner Vogels, recently asked: “Do you really need to have the biggest, highest-end model to solve this? The answer is no, you don’t.” You don’t send a surgeon to draw the blood. Match the model to the job, and health systems stop paying for capabilities they don’t need.
What Frontier Pricing Actually Costs
Frontier models are priced by the token, and healthcare runs this work at enormous volume. Serving a 7 billion-parameter model costs 10 to 30 times less, in latency, energy and raw compute, than running the same request through a 70 to 175 billion-parameter model, according to a 2025 Nvidia Research paper on agentic AI. A manual prior authorization already costs providers $12.88, dropping to $5.38 once it goes electronic, according to the 2024 CAQH Index, and the wrong model layered on top can erase those savings entirely.
Uber burned through its entire 2026 AI budget in four months, then capped what each engineer could spend. I’m hearing customers ask a new question now. What does it cost to own this model for the next three years, infrastructure and staffing included? That’s a total-cost-of-ownership conversation nobody was having two years ago.
I’m not arguing against frontier models. Health systems use them for good reasons. One contract and one API get you a model that takes on nearly any task, without months spent labeling data or staffing a dedicated team. When you’re trying to prove AI works, it’s the right call.
Frontier models also do things smaller models can’t. It can read a chart spanning a decade and three health systems, and reason through a case it has never seen. A model fine-tuned to sort clinical notes is good at exactly that, and not much else.
The kind of work will help determine the model you use. Rare or unfamiliar work belongs on a frontier model. The prior auths and coding queues that look identical 10,000 times a day don’t need one.
How Do You Own The Layer?
Owning the layer means using the right model for each task, on infrastructure you control.
Even without an AI model team, health systems can run a routing layer in front of their models, sending each task to the cheapest model already tested and shown to be accurate. It saves frontier models for problems only they can solve.
The radiology study mentioned earlier found that fine-tuning on 1,000 or fewer annotated reports pushed accuracy even higher than the zero-shot scores. A separate model, built on just 350 million parameters, ran on ordinary computers without a computing cluster or a data center, outperforming GPT-4o on medical report labeling.
Owning your models offers benefits beyond cost. You can fine-tune it as your data changes and run it yourself, so a vendor can’t retire the model or raise its price on you. People won’t use a system they don’t trust, and a model you can inspect yourself is easier to trust than an API you can’t.
Concentrating this much intelligence behind several closed providers carries real risk, however safe it looks today.
What It Takes To Scale
Health systems don’t need an AI lab, and nobody is asking them to walk away from frontier models. The systems that keep scaling treat cost and fit as an ongoing decision rather than something settled once during a demo, back when volume was small enough that nobody was watching the cost. The rest are still paying for a choice they never made.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
