Decision Models: Cut Your AI API Bill by 90%

If you run anything on AI today — a support bot, a content pipeline, an agent that routes tasks — you are almost certainly overpaying for intelligence you never use. Decision models fix exactly that. Instead of writing paragraphs, they return typed verdicts: a probability, a choice, a score. The first week of October 2026 brought a wave of them from Cloudflare, Amazon, OpenAI, Perplexity, and Microsoft, priced from four cents per million input tokens. If your pipeline spends frontier-model money on yes-or-no questions, this tutorial shows you how to stop.

What decision models are (and what they replace)

A decision model takes a state — text, JSON, an image, even a video frame — plus a schema of typed questions, and returns one verdict per question in a single pass. No tokens are generated. No parsing of free-form prose is needed. These are the jobs they cover: content moderation (safe or not?), ticket routing (billing, technical, or refunds?), classification (spam or legit?), scoring (how risky is this transaction?), and extraction triage (does this document contain an invoice?).

These tasks are verdicts, not essays. Yet most teams still route them through a general-purpose LLM, paying generation prices and waiting on autoregressive decoding for an answer that could have been one number. A decision model replaces that step with a single pass that scores the allowed options directly — faster, cheaper, and deterministic.

The October 2026 lineup: your options compared

Here are the decision models from the October 2026 wave, verified in the month’s model-release reporting:

Model	Price (input)	Deployment	Best fit
Cloudflare Clef 27B	$0.24 / 1M tokens on Workers AI	Cloud (Workers AI)	Multimodal verdicts at scale: text, JSON, images, video; 65,536-token context; post-trained from Qwen bases
Cloudflare Clef-flash 9B	$0.09 / 1M tokens on Workers AI	Cloud (Workers AI)	Ultra-fast hot-path decisions inside agents
OpenAI Decisions API	$0.10 / 1M input tokens, no output or cache charges	Cloud API (public beta)	Teams already on OpenAI: turns text/images into typed verdicts, built on Luna
Perplexity pplx-decider-v1-27b	$0.04 / 1M input tokens	Cloud API	Cheapest October launch (per Capital & Compute) for pure classification workloads
Amazon Strands Decider 2B	Free (open weights)	Local CPU or GPU	Privacy-sensitive or offline work; ships with weights, training data, and training scripts
Microsoft-Decision-1	$0 output pricing per million output tokens (per Capital & Compute)	Cloud API	Decision-heavy flows where output costs dominate
Inception Mercury Decide	See OpenRouter listing	Via OpenRouter	Teams that already buy inference through OpenRouter

The cost math: what 10 million calls really costs

Let’s make it concrete with an illustrative example. Suppose your product makes 10 million classification calls per month, each carrying roughly 1,000 input tokens. That’s 10 billion input tokens monthly:

At $0.04 / 1M input (Perplexity pplx-decider-v1-27b): $400/month.
At $0.24 / 1M input (Cloudflare Clef 27B): $2,400/month.
The same calls on a frontier model at several dollars per million input tokens climb into the tens of thousands per month — for answers that are still just yes/no verdicts.

Savings come from two directions: cheaper per-token rates and zero output tokens, because a decision model never generates a paragraph when a score will do.

Step-by-step: switching your classification calls

Step 1: Audit your current LLM calls

Export a week of API usage and flag every call whose output you reduce to a label, a boolean, a route, or a score. In most products this is a large share of total calls. Those are your candidates.

Step 2: Isolate the classification-style calls

Rewrite each candidate as a typed question: state in, options out. “Is this review fake? (yes/no).” “Which department? (billing/technical/refunds).” If a call can’t be rewritten this way, it stays on the general model.

Step 3: Pick a decision model from the lineup

Need multimodal input or long context? Clef 27B ($0.24/1M). Already on OpenAI? The Decisions API beta ($0.10/1M input, no output charges). Privacy constraints? Strands Decider 2B runs on your own CPU or GPU, free, with training data and scripts included.

Step 4: Run a parallel evaluation

Replay a few thousand real historical calls through both your current model and the decision model. Compare verdicts, not prose. Measure agreement rate and calibration — decision models return probabilities, so check that an 80% verdict is right about 80% of the time.

Step 5: Switch over, with a fallback

Route the classified calls to the decision model and keep the frontier model as fallback for low-confidence verdicts. You pay pennies for the routine cases and frontier rates only where judgment is genuinely needed.

When NOT to use decision models

Decision models are not a replacement for reasoning. Do not use them for open-ended generation (drafting, writing, creative work), reasoning-heavy work (multi-step math, complex analysis, coding), or any task where the answer is novel text rather than a choice among known options. They are the specialists in your stack, not the generalist.

What this means for you

If you’re a freelancer or small business running AI features, audit your API bill this week: the classification, moderation, and routing calls you’re paying generation prices for are exactly what the October decision-model wave was built to absorb. Swapping them to a $0.04–$0.24/1M-input verdict engine can cut that line item dramatically — often the difference between an AI feature that loses money and one that scales profitably. For developers, decision models belong on the hot path: a fast verdict before an expensive reasoning step means fewer wasted frontier calls. Start with your highest-volume classification call, evaluate in parallel, and switch.