Perplexity Decider v1.1: Open Model Halves Price, Tops Index

Two releases in one week

Perplexity has moved quickly with its new class of "decision" models. On October 1, 2026, the company opened a Decisions API backed by pplx-decider-v1-27b. The model is a 27-billion-parameter fine-tune of Qwen3.8-27B, released with open weights under Apache 2.0 on Hugging Face 45. Six days later, on October 7, it shipped pplx-decider-v1.1-27b. The update cut the hosted price in half and placed first on Hugging Face's newly launched Decision Index 0.3 13.

The pricing change is simple. At launch, the API charged $0.04 per million input tokens, with output free 45. Version 1.1 drops that to $0.02 per million input tokens, and output remains free 123. The model is also listed on OpenRouter, alongside Perplexity's own API 3.

What a "decider" actually does

These models do not produce prose. They pick from options you define and attach a probability to each one 1. A developer sends text, or text plus images, along with a typed schema. The schema describes the questions as yes/no, multiple choice, or an ordered score 45. The model returns calibrated probabilities per option, so there is no free-form output to parse or broken JSON to repair 4.

This is a clear design bet. Many production LLM workloads are really classification tasks in disguise: routing tickets, flagging content, judging whether a retrieved passage supports an answer. A model that skips generation and returns probabilities is cheaper to run and easier to trust in automated pipelines. Calibrated scores also let teams set their own thresholds instead of accepting a model's verbal confidence.

Perplexity's documentation reports response times under two seconds for a few hundred input tokens. That rises to about 23 seconds near the top of the context window 5. Reports differ slightly on that window. Some list it as roughly 250K tokens 12, while others give the exact figure of 262,144 tokens 45. Running the weights locally takes about 49GB of GPU memory 12.

Reading the benchmarks carefully

The headline claim is that v1.1 tops Hugging Face's Decision Index 0.3 with a score of 62.8. Jev 1.13.0, shown as a reference, scored 60.1 2. Perplexity's CEO went further on October 9, saying the model also leads the open DecisionBench benchmark at the lowest cost 3.

The earlier v1 numbers deserve a closer look. On Perplexity's own panel of 11 benchmarks and 7,210 rows, dated September 2026, pplx-decider scored 85.71%. Jev scored 84.51% and the base Qwen3.8-27B model scored 74.76% 25.

The per-benchmark breakdown is mixed:

• Biggest win: RAGTruth, at 88.80% versus Jev's 77.27% 25.
• Other strong result: FinancialPhraseBank sentiment, where pplx-decider also led 2.
• Where Jev led: WinoGrande, BBH and JevBench's hard public set 2.
• Effectively tied: JudgeBench 2.

One outlet put it plainly: this is a strong rival in the same tier, "not a coronation." It also noted that no independent rerun of the numbers exists yet 4. Another said it had not loaded the weights itself 2.

The overlap between sources is notable. Nearly every quantitative claim traces back to Perplexity's own chart or its Hugging Face model card. The Decision Index is also new and on version 0.3, so it has little track record.

Why it matters for open-source tooling

The RAGTruth result stands out more than the overall average. RAGTruth measures whether generated answers stay faithful to retrieved sources. An open, Apache-licensed model that does well there could slot into hallucination detection, content moderation, or policy-enforcement layers. Teams could run it on their own hardware rather than sending sensitive data to a hosted API.

The permissive license and self-hostable weights matter here as much as the price cut. Security-minded teams often avoid cloud classifiers for compliance reasons. A 49GB footprint is substantial but manageable on a single high-end accelerator.

The verdict

Perplexity has made a credible case that dedicated decision models are a useful category, and it is competing hard on price. Halving costs within a week suggests it wants developer adoption more than margin.

The performance story, however, is still mostly self-reported. The lead over Jev is narrow on aggregate and reverses on several individual tests. Teams should test v1.1 against their own classification workloads before treating its first-place ranking as settled. Independent reruns, which reportedly do not yet exist, will be the real test 4.