Perplexity's Decider V1.2 models now available on AI Gateway

Vercel has added Perplexity Decider V1.2 27B and Decider Flash V1.2 4B to AI Gateway, giving engineering teams access to specialized multimodal decision systems designed for classification, routing, and scoring. Rather than returning freeform conversational text, these specialized models evaluate text prompts, structured data records, and image inputs to answer typed questions with calibrated probability scores. The launch brings Perplexity decision architecture directly into Vercel infrastructure, according to Vercel Blog's report, enabling developers to evaluate complex incoming payloads against deterministic criteria without configuring custom inference pipelines.

This report breaks down the architectural capabilities, API integration patterns, operational cost structures, and practical evaluation workflows introduced with the Decider V1.2 family. It is written for backend engineers, system architects, and technical product leads who construct automated triage systems, customer support routing engines, and content moderation pipelines. Readers will learn how the two model tiers function across three distinct question types, how multimodal records route through the unified gateway, and how to execute decision queries directly within production TypeScript applications.

What shipped in Perplexity's Decider V1.2 models now

The release introduces two distinct model variants built specifically for deterministic judgment tasks rather than open-ended dialogue generation. Decider V1.2 27B, accessible under the model identifier perplexity/pplx-decider-v1.2, delivers twenty-seven billion parameters designed for nuanced semantic evaluations, complex organizational guidelines, and elaborate scoring rubrics. Alongside this flagship variant, Perplexity deployed Decider Flash V1.2 4B under the identifier perplexity/pplx-decider-flash-v1.2, offering a lightweight four-billion parameter footprint tuned specifically for high-throughput classification, rapid content filtering, and high-volume request routing where clear execution criteria and minimal latency are vital.

Both systems process multimodal payloads by inspecting plain text strings, structured JSON records, and base64 image objects embedded directly inside the state parameter. Instead of parsing fuzzy text output, client applications receive typed probability distributions across three standardized query schemas. Boolean queries calculate the exact statistical likelihood of a yes response, Choice queries select the optimal category from an application-defined list, and Score queries evaluate provided input against an ordered numeric rubric. Crucially, engineering teams can bundle multiple independent questions into a single request, allowing one input state to resolve multiple classification branches simultaneously.

What improved in Perplexity's Decider V1.2 models now

Operational efficiency improves dramatically by eliminating the fragile post-processing steps typically required when coercing general-purpose language models into returning reliable categorization data. By natively outputting calibrated probabilities alongside discrete decisions, these models allow systems to automatically triage borderline cases and divert uncertain scores directly to human reviewers. Organizations can replace sprawling prompt engineering chains and fragile regex extractors with structured calls that execute against predictable mathematical boundaries, whether identifying defective shipments, flagging policy violations, or sorting customer communications across support tiers.

Deployment economics also see immediate gains because Vercel routes inference traffic through AI Gateway using raw provider pricing without applying any platform markup or surcharge fees. Engineering teams configuring custom deployments can also leverage Bring Your Own Key credentials to preserve existing provider terms while gaining full platform visibility. Requests automatically register inside centralized spend budgets and searchable telemetry logs, allowing teams to monitor usage, track unit economics, and prevent cost overruns without instrumenting separate monitoring infrastructure across disparate cloud services.

What you gain from Perplexity's Decider V1.2 models now

Developers integrating these models gain unified access to three versatile consumption pathways comprising the official AI SDK, the dedicated Decisions API, and the TypeSafe-compatible API interface. This flexibility ensures that development teams can implement decision logic using their preferred architectural paradigms while maintaining type safety across the entire application stack. Furthermore, developers can immediately test and calibrate prompts inside the hosted model playground for both Decider V1.2 27B and Decider Flash V1.2 4B, validating complex rubrics before writing a single line of backend integration code.

Applications also gain robust multimodal processing capabilities through standard JSON payloads without requiring specialized pipeline middleware or external image conversion services. By embedding base64 image_url objects directly into the JSON state object, a single request can assess visual damage alongside accompanying textual incident descriptions. A logistics platform, for example, can determine whether a package arrived damaged, score the severity of the damage against established insurance criteria, and route the customer ticket to a priority queue in a single network round-trip.

How to get Perplexity's Decider V1.2 models now

Getting started requires installing the Vercel AI SDK into your modern JavaScript or TypeScript backend project and configuring your AI Gateway credentials. The gateway enables teams to authenticate against multiple model providers using a single centralized API key, eliminating secret sprawl across deployment environments. After installing the core package dependencies, developers invoke the experimental_decide function by importing it directly into application routing handlers, background worker jobs, or serverless functions to process incoming payload state objects against defined evaluation questions.

npm install ai

```javascript
import { experimental_decide as decide } from 'ai';

Switching between the high-capacity twenty-seven-billion parameter model and the rapid four-billion parameter flash variant requires only updating the model string parameter. To route high-throughput workloads to the smaller model, developers set the model parameter to perplexity/pplx-decider-flash-v1.2 inside the configuration object. Teams can explore the complete decision model catalog within the dashboard, examine live telemetry in request logs, and test multimodal payloads containing base64 image data directly within the interactive model playground.

What to watch after Perplexity's Decider V1.2 models now

As teams deploy decision models into critical production flows, system operators must actively monitor probability distributions across borderline cases and set clear confidence thresholds for automated actions. Because boolean, choice, and score queries deliver statistical likelihoods rather than absolute binary answers, engineering teams should establish automated escalations that flag uncertain decisions for secondary human verification. Configuring these routing safeguards ensures that edge cases, ambiguous customer tickets, or unclear product images are reviewed by human agents before downstream fulfillment or refund actions execute.

Teams should also track gateway analytics and request logs to observe how multimodal payload sizes influence latency across different geographic regions and compute tiers. Because image inputs rely on base64 image_url objects nested in JSON state payloads, larger graphic assets will increase network payload transfer sizes and request processing duration. Monitoring overall usage against centralized spend budgets and exploring the broader decision model catalog will allow infrastructure teams to optimize cost and performance as Perplexity continues developing specialized decision architectures.