Models & labs

Perplexity Decider v1.3: Open-Weight Decision Models Learn Text Extraction, Undercutting Gemini Decision 1 Fivefold

Perplexity is updating its decision models, which output probabilities rather than text. Decider v1.3 (27 billion parameters, based on Qwen3.8-27B) and Decider Flash v1.3 (based on Qwen3.5-4B) are released under an Apache 2.0 license and introduce Locate, an extraction feature for exact text spans. Via the API, they cost $0.02 and $0.01 per million input tokens—five to ten times less than Gemini Decision 1.

Abstract illustration: a lavender light beam passes through a glass block etched with lines and isolates a few glowing fragments.
AI-generated illustration

Decision models, which answer closed-ended questions with probabilities instead of writing out a response, have become a distinct category in just a few weeks. Google just entered the space with Gemini Decision 1. Perplexity is responding with version 1.3 of its Decider models: still open weights, still the same pricing, and now capable of extracting text spans.

Two Models, One New Feature

The family comes in two sizes. pplx-decider-v1.3 is fine-tuned from Qwen3.8-27B and requires a GPU with roughly 49 GiB of memory for its weights; pplx-decider-flash-v1.3, distilled from the larger model onto a Qwen3.5-4B base, fits into roughly 8.5 GiB. Both are released under an Apache 2.0 license, which permits commercial use, and the repository is public: we were able to download the files without an access token on November 14, despite the card still claiming otherwise.

Like previous releases, they answer multiple-choice questions (up to 255 options), yes/no prompts, or rating scales, using text or image inputs. The newcomer is Locate: rather than selecting from predefined options, the model returns passages from the text matching an instruction (“error codes,” “the person who signed the contract”), copied verbatim along with their exact character offsets and a confidence score. Intended use cases include named-entity recognition (NER), PII detection, and extractive question answering. Locate works only on text, handling up to 128 questions per request. It relies on a separate adapter loaded on demand, and mixing Locate with decision questions in a single request adds latency.

Pricing Compared to Google

In Perplexity’s Decisions API, v1.3 costs $0.02 per million input tokens and Flash costs $0.01, with no per-request fees and no billed output tokens—unchanged from v1.2 pricing. Gemini Decision 1 is priced at $0.10 per million input tokens, making it five times more expensive than Decider v1.3 and ten times more than Flash. And this cost advantage does not come at the expense of quality. Artificial Analysis, which has already included v1.3 in its Decision Index v2.1 (page accessed on November 14), awards it 70 points, compared to 66 for Gemini Decision 1, with a measured cost of roughly $68 per million tasks versus $345. Flash v1.3 gained four points over its predecessor (64 versus 60) for roughly $34 per million tasks. The larger model barely advanced over v1.2 on this index and remains behind the leading trio at 73 points (Drex Max v1.1, SubQ Snap 1.2, and GLiDE v1.2). The index does not evaluate Locate.

Our take

With Locate, Perplexity expands its decision models beyond simple sorting: a single API call can now classify a message, score its urgency, and extract relevant details—three tasks that previously required a generative model or a specialized extractor. Facing Google, the argument is no longer just the five- to ten-fold price advantage, but open access: with Apache 2.0 weights, a company can run the Flash model on a single mainstream GPU and keep its data on-premise, a critical factor for PII detection. It remains to be seen how well the extraction holds up on real-world internal documents, as the Artificial Analysis index does not evaluate Locate; Perplexity itself advises calibrating probabilities against actual production traffic. Furthermore, the 27B model shows little progress in pure decision accuracy: the highlight of this release is unequivocally Locate.

Sources

  1. Hugging Face — perplexity-ai/pplx-decider-v1.3 (model card, Apache 2.0 license) Primary source
  2. Hugging Face — perplexity-ai/pplx-decider-flash-v1.3 (model card) Primary source
  3. Perplexity Docs — Pricing (Decisions API) Primary source
  4. Perplexity Docs — Decisions API Quickstart Primary source
  5. Google — Gemini Decisions API: Fast, calibrated multimodal decisions (November 11, 2026) Primary source
  6. Google AI Studio — Gemini Decision 1 model card Primary source
  7. Artificial Analysis — Decision Index v2 (results page) Secondary source
  8. Artificial Analysis on X — Gemini Decision 1 cost per million tasks (November 11, 2026) Secondary source

A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.