Cheap, fast models make a whole class of AI products viable

Calvin French-Owen makes a simple economic argument in "Small Models Have Arrived": the recent wave of small, fast models has quietly changed which products are worth building. His example is concrete. A personalized news task that used to cost about a dollar to run on older models now costs around ten cents with gpt-5.6-luna, which serves roughly 100 tokens per second. That is a tenfold drop, and it moves consumer AI features from "too expensive to ship" to "obviously worth it."

Drawing on a conversation with his co-founder Peter Rein, French-Owen splits work into two kinds. There is "IQ 180" work, the novel problem-solving that needs a genius-level insight, and there is "token spewer" work, steady incremental progress across many small fronts. He estimates about 95% of business work is the second kind, where responsiveness and cost matter far more than raw brilliance. Frontier models still earn their price on the hard 5%, breakthroughs in engineering and science, but for the rest he expects demand to shift toward fast, cheap, good-enough models. He also notes GLM 5.3 as a new contender on the price-performance frontier, a reminder that this is a competitive race down the cost curve, not a single vendor's story.

Why it matters

If you are deciding which model to put behind a feature, the default of reaching for the strongest one is now often the wrong call. Profiling your workload into the 95% that a small model handles and the 5% that genuinely needs a frontier model can cut your inference bill by an order of magnitude, and it makes products that were uneconomical last year worth prototyping today.