Perplexity Decider v1.3 makes AI routing a replaceable infrastructure layer
An open-weight model for classification, routing and text extraction changes how teams should build the control plane around expensive generative models.
What shipped
Perplexity released Decider v1.3 and the smaller Decider Flash v1.3 on November 12. Instead of generating prose, the models return probabilities for yes-or-no decisions, fixed choices and rubric scores. A new Locate function extracts exact spans from source text with offsets and confidence scores.
Both models are available through the Perplexity Decisions API, OpenRouter and as Apache-2.0 weights. Official API pricing is $0.02 per million input tokens for the larger model and $0.01 for Flash, with no output-token charge.
The control plane
This is not a replacement for a model that writes, codes or plans across multiple steps. It is a specialised control layer for cases where software needs an actionable judgment: request routing, tool selection, moderation, policy checks, result ranking or field extraction.
The operational change is that teams can evaluate several bounded questions over shared input without generating and parsing text. Decisions become easier to measure, thresholds can change without rewriting prompts, and outputs are cleaner to audit.
What we would do
We would begin with one frequent decision currently handled by a generative model. We would build a labelled set from real traffic, compare errors, latency and cost per completed task, then run the new model in shadow mode before letting it affect users.
We would keep the interface provider-neutral and retain human review for ambiguous or consequential cases. Self-hosting would come later, only when privacy, volume or vendor lock-in justifies the GPU footprint.
Perplexity’s Decider v1.3 is not the model we would put at the end of an agent; it is the model we would test at the beginning. Too many production systems pay a generative model to answer a bounded question — route this ticket, approve this tool, score this result or find this field — then parse prose and hope the answer is stable. Decider v1.3 and Flash v1.3 return probabilities, add exact-span Locate extraction, ship through an API and publish Apache-2.0 weights, making the decision layer easier to swap and own. Our move would be to place a provider-neutral decision interface ahead of expensive reasoning models, replay real traffic, calibrate thresholds on labelled examples and shadow-route before changing production. We would not replace complex reasoning or high-stakes review, and we would not treat Perplexity’s own benchmarks as a purchasing decision. The Hugging Face card also says the 27B model needs about 49 GiB for BF16 weights plus working memory and requires authenticated access, so open weight does not mean effortless self-hosting. Still, this is worth touching now: the likely win is not a smarter chatbot, but lower cost per completed task, fewer parsing failures and less lock-in in the control plane.
Turn insight into a working system.
We build a provider-neutral decision layer that classifies, routes and extracts before an agent spends on a generative model. Virtual Arc adds representative evals, shadow routing, confidence thresholds, permissions and audit trails. Talk to us if your system uses an LLM mainly to return labels or routing choices.