Not every notable AI release this year has been a bigger chatbot chasing a higher benchmark score. On October 1, Perplexity opened a Decisions API running on a model that was never meant to hold a conversation at all. pplx-decider-v1-27b doesn't write sentences — it takes a question and a set of candidate answers you define, and returns a probability for each one, in under 50 milliseconds, at $0.04 per million input tokens with output tokens free.
That's a narrow job, but it's an enormous one in practice. Every AI product that classifies, moderates, routes, or grades content behind the scenes is currently paying a large language model to do a job that doesn't need full text generation at all — and Decider is built specifically to replace that waste.
Quick summary: Perplexity opened its Decisions API on October 1, 2026, powered by pplx-decider-v1-27b — a 27-billion-parameter, open-weight model on Hugging Face under Apache 2.0, fine-tuned from Qwen3.8-27B. It's built for classification, routing, content moderation, and rubric grading: a request can carry up to 128 questions about the same content, and the model returns normalized probability distributions across user-defined candidate classes rather than free-form text. Pricing at launch was $0.04 per million input tokens with output tokens free (per Perplexity CEO Aravind Srinivas), with sub-50ms response times — cut further to $0.02 per million input tokens when Perplexity shipped v1.1 on October 6. On Perplexity's own 11-benchmark, 7,210-sample evaluation panel, Decider scored 85.71% against 84.51% for TypeSafe's Jev — the browser-agent decision model LumiChats covered previously — with the widest gap on the RAGTruth benchmark, 88.80% versus 77.27%.
What Decider Actually Does
Most production AI systems quietly rely on a large language model for jobs that are really classification problems wearing a chat interface: is this message spam, which of these five intents did the user mean, does this generated answer actually match the source document. Decider is built to answer exactly that kind of question directly. Instead of generating a text reply that application code then has to parse, it outputs a normalized probability distribution across whatever candidate classes you define for that request — a structural change that's faster and more reliable for anything downstream that needs a clean, thresholdable number rather than a sentence to interpret (Source: systemonemodels.org).
It's a 27-billion-parameter model, fine-tuned from Alibaba's Qwen3.8-27B, released open-weight on Hugging Face under the Apache 2.0 license — meaning it can be self-hosted and modified freely, not just accessed through Perplexity's own API. A single request can bundle up to 128 separate questions about the same piece of content, useful for a moderation pipeline that needs to check one document against many policy rules at once.
How It Stacks Up Against Jev
The most direct comparison Perplexity itself draws is against Jev, the browser-automation decision model from TypeSafe that LumiChats covered in a recap of mid-September's AI news. On Perplexity's own 11-benchmark, 7,210-sample evaluation panel, Decider scored an overall 85.71% against Jev's 84.51% — a real but modest edge. The gap widens considerably on one specific benchmark, RAGTruth, which tests whether a model's output stays faithful to its source material: Decider hit 88.80% versus Jev's 77.27% (Source: aiweekly).
| Perplexity Decider v1 (27B) | TypeSafe Jev | |
|---|---|---|
| What it's for | Classification, routing, moderation, grading | Browser-agent action selection |
| 11-benchmark panel score | 85.71% | 84.51% |
| RAGTruth score | 88.80% | 77.27% |
| License / access | Open-weight, Apache 2.0, plus paid API | Proprietary — hosted API only, no published weights (an open-source ecosystem of wrapper tools exists around it, e.g. Browser Use's agent) |
These benchmark numbers come from Perplexity's own evaluation panel, not an independent third party — a real edge over Jev is plausible given the task-specific design, but treat the exact percentages as Perplexity's own measurement until someone outside the company reproduces them.
Why This Matters Even If You'll Never Call the API Yourself
Models like Decider rarely make headlines the way a new flagship chatbot does, but they're a meaningful part of why AI products you already use get cheaper and faster over time. A search engine's content filters, a chat app's intent router, a customer-support tool's routing logic — all of that infrastructure has historically run on a general-purpose LLM that's slower and more expensive than the job actually requires. Purpose-built, open-weight classifiers priced in fractions of a cent per million tokens are the kind of unglamorous efficiency gain that eventually shows up as a lower subscription price or a faster response, even if you never see the model's name.
01What is Perplexity Decider v1, in plain terms?
It's a small, open-weight AI model built to classify, route, moderate, or grade content — answering structured questions with probabilities rather than holding a conversation. It's meant to replace larger, slower chat-style models that are currently doing that same narrow job less efficiently.
02Is Decider v1 free to use?
The model weights are open on Hugging Face under the Apache 2.0 license, so you can self-host it for free. Using Perplexity's own hosted Decisions API costs $0.04 per million input tokens, with output tokens free.
03How does it compare to Jev?
On Perplexity's own 11-benchmark panel, Decider scored 85.71% versus Jev's 84.51%, with the largest gap on the RAGTruth faithfulness benchmark. The two models aren't built for identical jobs, though — Jev is specifically for browser-agent action selection, while Decider is a general-purpose classifier.
04Do I need this if I'm not building AI infrastructure myself?
No — this is a developer-facing, backend model. It matters to everyday users indirectly, as one of the pieces that makes AI products faster and cheaper to run.
If you're more interested in which AI model actually writes, codes, or reasons best for your own work than in backend infrastructure models like this one, LumiChats gives you one subscription with 40+ AI models — including Perplexity, Claude, GPT, and Gemini — so you can compare them directly instead of reading spec sheets.
