Perplexity's Decider V1.2 Models Are Now on Vercel AI Gateway, and They Think in Probabilities

Perplexity's Decider V1.2 27B and Decider Flash V1.2 4B are now live on Vercel's AI Gateway. These are not chat models. They are decision models, a specific type of AI built to classify, score, and route content rather than generate text, and that distinction is worth understanding.

So What Is a "Decision Model" Exactly?

Most AI models you've played with are generative: you ask a question, they write something back. Decision models work differently. Instead of producing prose, they answer structured questions with probabilities and categories.

Concretely, the Decider models handle three types of questions:

• Boolean questions return a likelihood of yes or no. ("Is this review spam?")
• Choice questions pick from a list of categories you define. ("Which support department should handle this ticket?")
• Score questions rate content against a rubric you write. ("How urgent is this complaint, on a scale you specify?")

The model doesn't just pick an answer. It returns the probability behind it, so your app can decide what to do with low-confidence results. If the model is only 55% sure a photo shows a damaged product, you can send that case to a human reviewer automatically. That kind of fallback logic is genuinely hard to build without a tool like this.

You can also run multiple questions against the same piece of content in a single request, which keeps things fast and cost-efficient.

Two Models, Two Use Cases

Perplexity ships two sizes here, and choosing between them is straightforward.

Decider V1.2 27B is the larger model. The "27B" refers to the number of parameters it was trained on, which you can roughly think of as its capacity for nuance. Use this one when your classification criteria are complex or your rubric has a lot of detail, like grading a portfolio submission or flagging subtle policy violations.

Decider Flash V1.2 4B is the smaller, faster sibling. It's designed for high-volume situations where the criteria are clear and you need speed: filtering thousands of uploaded images, tagging support tickets at scale, or labeling form submissions in real time.

Both models are multimodal, meaning they can process images alongside text. You can send a photo of a product return, a screenshot of a review, or a user-uploaded image and ask the model to make a call on it.

How a Designer-Builder Could Actually Use This

Say you're building a community app where users submit work for feedback. Right now, moderating that flow probably means you do it manually or you leave it unmoderated and hope for the best.

With the Decider models through AI Gateway, you could set up a simple routing layer. Every submission runs through a Choice question: is this appropriate for the main feed, should it go to a review queue, or should it be flagged for moderation? You write the categories. The model scores each submission against them. Borderline results, anything below a confidence threshold you choose, go to a human. Clear-cut cases move automatically.

This is the kind of logic that used to require a dedicated back-end engineer and a hand-rolled classifier. Now it's accessible through the same AI Gateway you might already be using to call other models, with one API key covering everything.

On the practical side, Vercel AI Gateway logs every decision request, tracks your spend, and lets you set budget limits. Pricing passes through at whatever Perplexity charges, with no platform markup. If you bring your own Perplexity credentials, that applies too.

The model playground on Vercel lets you test Decider V1.2 27B before you wire anything up. That's a good place to start: write a few test questions, throw some real content at it, and see whether the outputs match your instincts before you build a workflow around them.

What to Keep in Mind

Decision models are not magic filters. They are only as good as the questions and rubrics you write for them. Vague criteria will produce inconsistent results, so invest time upfront in writing precise, testable questions. Think of it less like prompting a chatbot and more like writing a brief for a junior reviewer: the clearer your instructions, the more reliable the output.

It is also worth noting that this is V1.2, which implies earlier versions exist and future ones will follow. The behavior may improve or shift over time, so if you build a classification workflow on this, keeping an eye on version updates is sensible practice.

The bigger picture here is not really about these two specific models. It is about a pattern: structured AI judgments, embedded directly into your product's logic, without needing a dedicated data science team. For designers who are learning to build, that pattern is one of the more genuinely useful things to get comfortable with.