The Return of the Specialist Model
As frontier-model API costs and latency have become a real budget line for AI-heavy products, more engineering teams are pulling narrow, high-volume subtasks — intent classification, entity extraction, routing decisions — off the frontier model and onto small, fine-tuned models trained for that one job.
Why This Wasn't the Default Already
Fine-tuning a small model used to require enough labeled data and ML tooling that only large teams could justify it. Cheaper, faster fine-tuning pipelines and the ability to bootstrap training data from a frontier model's own outputs have lowered that bar considerably.
The Resulting Architecture
The pattern taking shape looks less like a single model doing everything and more like a large general model handling reasoning and generation, orchestrating calls out to a set of small specialist models for the repetitive, well-bounded pieces of the pipeline — trading a bit of engineering complexity for a meaningful cut in per-request cost.