Text Classifiers: Local Models vs Always-On Endpoints

Daniel has a home inventory system, a box of soldering supplies, and one question: is there a casual SaaS for text classification, or should he just run a private model locally? The answer turns out to be a landscape, not a product. Specialist classifiers beat general LLMs on both cost and accuracy — roughly twenty-five times cheaper and ahead on intent benchmarks — which makes the local instinct right and the SaaS pitch harder to justify. The features Daniel wants do exist, several times over: upload labeled pairs, retrain on a cadence, call inference over a stable endpoint. What doesn't exist is a household name for it. Meanwhile AWS Comprehend charges for a provisioned endpoint whether or not anything is flowing through it, and a six-year-old forum post asking for exactly Daniel's use case got an answer that still applies today. Key sources: • Amazon Comprehend Pricing — https://aws.amazon.com/comprehend/pricing/ • HF Inference Providers, Text Classification task — https://huggingface.co/docs/inference-providers/en/tasks/text-classification • HF Inference Providers overview — https://huggingface.co/docs/inference-providers/en/index Episode 5804 — myweirdprompts.com/n/5804 (permanent short link: myweirdprompts.com/725406)