The question
What we told them
Instinct is wrong here, and the industry's own results say so.
Models go through two kinds of schooling. First they read an enormous amount of text and gain raw capability. Then they're tuned — taught to follow instructions, and in a final stage trained on thousands of human judgments about which of two answers is better. That last stage is what turns raw capability into usefulness.
How much does it matter? In the study that established the method, human raters preferred a 1.3-billion-parameter tuned model over an untuned model roughly a hundred times larger. A year later, a 7-billion-parameter open model tuned with a newer preference method beat the strongest 70-billion chat model of its day on standard benchmarks. Tuning beat a hundredfold size advantage in one case and a tenfold in the other.
For your decision, that means two things. First, "how big" is the wrong axis. Ask each vendor what the model has been tuned and evaluated to do — on what data, with what checks around it. A competent mid-size model with the right context, tools, and verification beats a giant generic one at a specific job, and it costs less on every request, forever. Second, favor use cases where output is checkable: code that has to run, extractions you can compare to the source, numbers that must reconcile. The newest training methods lean on exactly that kind of automatic verification, and it's also where you can audit results yourself.
The biggest frontier models earn their premium on genuinely hard, open-ended reasoning. Most business workflows aren't that.
Ask what the model was tuned to do, not how big it is. A well-tuned smaller model routinely beats a giant generic one at a specific job — at a fraction of the per-request cost.