RESEARCH · MODEL SELECTION
Big model,
small model.
The reflex is to reach for the biggest model and never think again. It is also how budgets and latency quietly fall apart. On most of the work a business actually runs, a smaller or open model matches the frontier at a fraction of the cost. The skill is knowing which tasks are the exception.
Route by measured difficulty: most traffic never needs the frontier.
THE REFLEX
The frontier is not always the answer.
A frontier model is a general-purpose instrument tuned for the hardest, most ambiguous problems. Most enterprise tasks are not that.
Classifying a ticket, extracting fields from an invoice, drafting a standard reply: these are narrow, high-volume, and forgiving. A small or open model handles them well. Paying frontier prices for them is paying for capability you never use, on every single call, forever.
WHERE SMALL WINS
Cheaper, faster, yours.
High volume, narrow scope.
When the same task runs a million times, the per-call price dominates everything, and a small model's savings compound into the real number on the invoice.
Latency-bound work.
A small model answers in a fraction of the time. For anything a person waits on, speed is not a nicety. It is the quality.
Privacy and control.
An open model you host keeps the data in your building, and can be fine-tuned on your domain until it beats a larger general model at your specific job.
WHERE BIG EARNS IT
Spend capability where it changes the outcome.
Reach for the frontier when the task is genuinely hard: ambiguous instructions, reasoning across many steps, long-horizon agency, a high cost of being wrong. On those, the gap between a frontier model and a small one is not a few points. It is the difference between done and not done.
The art is spending that capability only where it moves the result, and not on the routine majority that a cheaper model would have finished just as well.
THE DESIGN
Route, do not standardize.
The mistake is choosing one model for everything. The design is a router: send the easy, high-volume tasks to the small model, escalate the hard ones to the frontier, and measure the crossover so the line sits in the right place for your work. Most teams discover the majority of their traffic never needed the big model at all.
HOW MILLIE MEASURES THIS
For every workflow we score the full field, small and open models included, on your tasks. The output is the crossover point: the exact task difficulty where paying for the frontier starts to pay off. Below it, the cheaper model wins on value.
See the frontier, priced for the public curve, and what we measure for how we re-draw it on your work.