The Quiet Story of October: Tiny Models That Just Decide
Perplexity, Cloudflare, Microsoft, and AWS all shipped decision models. Cheap, local, and boring in the best way.
While everyone watched the big chatbots, a different category shipped: decision models that return a choice, not a paragraph.
Perplexity's pplx-decider-v1-27b runs at about four cents per million input tokens with free output. Cloudflare's Clef and AWS's Strands Decider do similar jobs you can self-host.
The point isn't cleverness. It's cost and latency for the boring calls: route this ticket, classify this input, pick this action.
For builders, this is the more useful release than another flagship. A four-cent decision beats a two-dollar essay when you make a million of them.
Expect more of this. The commodity layer of AI is getting cheap fast, and that's where products get built.
Updated October 28, 2026: the category below has only grown since - every major platform now ships one.
Frequently asked questions:
Q: What can decision models do?
A: The boring calls: route this ticket, classify this input, pick this action. They return a choice, not a paragraph.
Q: How much do they cost?
A: Perplexity's pplx-decider-v1-27b runs about four cents per million input tokens with free output; Cloudflare's Clef and AWS's Strands Decider are similar and self-hostable.
Q: Why are they a bigger deal than another flagship?
A: A four-cent decision beats a two-dollar essay when you make a million of them. The commodity layer is where products get built.