The Decision-Model Gold Rush: Why Everyone Shipped a Tiny Model This Month
A new category showed up in October and every lab raced to fill it. The prices are absurd.
Something odd happened in early October. Instead of bigger chat models, labs shipped tiny 'decision' models that take your state and a typed question and return a choice, a score, or a probability in one pass.
OpenAI put GPT-6 Luna behind a Decisions API on October 6, promising answers about ten times faster than the Responses API at $0.10 per million input tokens with no output charge. Cloudflare followed with Clef, Perplexity with pplx-decider at $0.04, Microsoft with Decision-1 at $0.042, and Liquid AI with d1 models that answer in eight milliseconds on a consumer GPU.
The appeal is simple: most 'AI' calls in a product are a classification in disguise. You don't need a reasoning model to decide whether an email is spam or which tier a user belongs to. A two-billion-parameter model does it for a fraction of a cent.
AWS even open-sourced Strands Decider 2B, so you can run the whole thing on your own hardware. The commodity wave here is real and it's moving faster than the flagship race.
If you're still prompting a frontier model to make a yes-no call a thousand times a day, this month's releases are a direct tax refund.
Updated October 28, 2026: the prices below were verified at announcement - several have fallen further since.
Frequently asked questions:
Q: What is a decision model?
A: A tiny model that takes your state and a typed question and returns a choice, score or probability in one pass - not a paragraph.
Q: How cheap are they?
A: OpenAI's Decisions API runs about ten times faster than Responses at $0.10 per million input tokens with no output charge; Perplexity's is $0.04; Liquid's d1 answers in eight milliseconds on a consumer GPU.
Q: When should I use one instead of a frontier model?
A: Whenever the task is a classification in disguise - spam or not, which tier, route or approve. A two-billion-parameter model does it for a fraction of a cent.