Perplexity has published a detailed post on Decider v1.3, the newest version of its decision model family, titled "Perplexity Decider v1.3: Training, eval, and use cases." The post hit the front page of Hacker News on Friday and drew more than 140 points within hours, with developers trading benchmark results and arguing over where these models actually fit. For more on this and other breaking AI news, we follow every model release as it lands.
The attention is less about one update and more about what the thread reveals: in barely a year, decision models have gone from a curiosity to a crowded category with their own benchmarks, their own pricing logic and a competitive field that now includes Perplexity, TypeSafe, Microsoft, Google, Cloudflare and OpenAI.
What a Decision Model Actually Does
The clearest definition comes from Artificial Analysis, the independent evaluation lab whose Decision Index v2.1 has become the reference leaderboard for the category. In its explainer, a decision model takes a question with a fixed set of options — for example, a support ticket asking whether to route to billing or technical, whether it is urgent, and how upset the customer is — and returns an answer for every option with a probability attached, all in a single call.
A general-purpose language model, by contrast, generates a structured output that then has to be parsed, and a reasoning model spends extra compute thinking before it answers. Decision models skip that. They are trained to choose, not to explain, which is what makes them fast and cheap enough to run on every document, every request, at high volume.
How the Category Is Measured
Artificial Analysis Decision Index v2.1, updated with a public methodology page, combines six evaluations with published weights: AA-Paperclip at 30 percent, which runs 1,500 resumes, receipts, invoices and research papers through decisions and fact extraction; EvalSafe-ONET-AA at 15 percent, based on TypeSafe's ONET benchmark of judgements about workplace documents; When2Call at 10 percent, NVIDIA's tool-calling evaluation; AA-Omniscience-Decision v1.1 at 15 percent for knowledge and hallucination; HLE-MultiChoice v1.1 at 15 percent, the text-only multiple-choice questions from Humanity's Last Exam with an option to decline; and AA-Symmetry v1.1 at 15 percent, which checks that a decision holds when its options are moved, reworded or negated.
That last test is the tell. A model that flips its answer when you reverse the order of options is not actually deciding — it is pattern-matching. Calibration, invariance and knowing when to decline are treated as first-class requirements, not bonuses.
What the Hacker News Thread Adds
Perplexity first open-weighted a 27-billion-parameter member of the family, pplx-decider-v1-27b, on Hugging Face in early October, so the community has had hands on these models for weeks. The reaction to v1.3 was mixed in a way that reads like a healthy market rather than a hype cycle.
One developer reported that Decider v1.3 showed gains over both v1.2 and TypeSafe's Jev 1.14 on their internal evals for an entity extraction and matching task. Another said it performs well on their local benchmark and is priced competitively, noting that a bidirectional-attention architecture may suit decision work. Others pushed back: one commenter found v1.3 a step back for their use case and said the public index agreed, and another questioned whether flash variants dilute the category's core promise of cheap, accurate bulk decisions.
The thread also surfaced the surrounding ecosystem — Cloudflare's Clef decision models and Clairic extraction model, the ScaleDown API, Liquid's D1 and OpenAI's own decision model — with several developers noting that Artificial Analysis' index, which includes cost per task and calibration alongside accuracy, is the only cross-vendor comparison they trust.
Why It Matters
We have covered this category's rise step by step: TypeSafe's Jev system-one model, Microsoft's Decision-1 built on Qwen, and Google's Decisions API with its Gemini Decision-1 model. Perplexity shipping v1.3 confirms the pattern — every major AI platform now wants to own the layer where structured choices get made, because that layer sits in front of everything: support triage, document processing, tool calling, moderation, compliance checks.
The economics are the point. A reasoning model can out-answer a decision model on accuracy, but at high volume its latency and cost make it the wrong tool for millions of daily classifications. Decision models trade general intelligence for calibrated, invariant, single-call answers, and as the Hacker News debate shows, buyers are now grading that trade with real benchmarks and real workloads.
Decider v1.3 will live or die on those scorecards. But the category itself has clearly arrived.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
