Perplexity Releases Open-Weight Multimodal Decision Models pplx-decider-v1.2 (27B) and Flash (4B) on OpenRouter and Vercel AI Gateway

Background and the Problem

In production agentic pipelines—spanning browser automation, multimodal customer triage, and coding agent tool selection—the majority of intermediate steps require discrete classification, routing, or policy verification rather than open-ended text generation. Using heavy autoregressive LLMs for every branching decision introduces unnecessary latency, JSON formatting failures, and inflated token costs.

Architecture and How It Works

Perplexity (@perplexitydevs) rolled out its next-generation open-weight multimodal decision model family, Decider v1.2:

1. Multimodal Input to Typed Probability Output: Both models accept text and image inputs and directly output structured, typed decisions accompanied by calibrated probability distributions;
2. Dual 27B and 4B Flash Checkpoints: pplx-decider-v1.2 upgrades the flagship 27B checkpoint succeeding v1.1, while pplx-decider-flash-v1.2 delivers a compact 4B architecture engineered for ultra-low latency at half the cost;
3. Unified Gateway Aliases: Supported natively across OpenRouter and Vercel AI Gateway with rolling aliases (pplx-decider-latest and pplx-decider-flash-latest).

Benchmarks and Measured Results

Key deployment metrics and pricing highlights:

• $0.02 per Million Input Tokens: The 27B pplx-decider-v1.2 model is priced at just $0.02/1M input tokens, with the 4B Flash model cutting inference cost in half again;
• Day-Zero Cloud Gateway Availability: Live immediately on OpenRouter (openrouter.ai/perplexity/pplx-decider-v1.2) and Vercel AI Gateway.

Getting Started for Developers

Integration guide for agent builders:

• Fast Tool & Guardrail Routing: Place pplx-decider-flash-v1.2 at the front of your agent loop to evaluate tool choices and visual UI states, escalating to frontier reasoning models only when the returned top-choice probability falls below a calibrated confidence threshold.