AI News
AI News AgentModel releasePerplexity4 min read

Perplexity launches Decider v1.3 for AI decision-making

Perplexity introduces Decider v1.3 and Decider Flash v1.3, models for classifying, scoring, routing queries, and extracting data from text and images. The new `locate` function identifies specific spans and returns their position, while the company reports improvements in accuracy, stability, and cost compared with several reference models.

Perplexity launched Decider v1.3 and Decider Flash v1.3, two models designed to make structured decisions instead of generating long answers. They can classify text and images, score results, decide which model should handle a query, and extract specific information from a document.

The distinction matters because many tasks currently sent to a generative model do not require writing. If you want to know whether a page is spam, whether a query requires web search, or whether a summary accurately reflects a document, you only need a decision and its probabilities.

A model for classification, scoring, and extraction

Decider accepts an input and one or more questions, each with its possible answers. It can handle:

  • Binary decisions, such as yes or no.
  • Multiple-choice classification, with custom labels.
  • Ordered scores, such as low, medium, or high.
  • Extraction of specific text spans.

The new locate function identifies exactly where the requested information appears. For example, it can identify a person's name, an email address, the city listed for a job, or the requirements of a posting, and return the position of those spans within the text.

It can also combine tasks. Given a job posting, the model could decide whether senior-level experience is required and extract the required skills at the same time.

Decider v1.3 is the main model. Decider Flash v1.3 is a smaller version created through distillation, a technique that transfers part of a large model's behavior to another model that is faster and cheaper.

More stable responses

Perplexity modified the architecture so the model can read the full context in both directions. In its tests, this bidirectional attention improved performance compared with the traditional operation of generative models, which process each token by looking only backward.

The company also separated questions asked about the same input more effectively. As a result, the answer to one question is less sensitive to adding or removing other questions in the same request.

In a test with 372 multiple-choice questions, Decider v1.3 changed its answer in only 0.27% of cases when the order, labels, or names of the options were modified. Perplexity reports rates between 2.69% and 4.84% for Jev 1.14, another decision model used as a reference.

Results reported by Perplexity

The figures come from internal and external evaluations described by the company. Across 63 query-classification tasks, Decider v1.3 reached an average accuracy of 91.6%, 6.54 percentage points above the classifiers Perplexity was using in production.

Across 72 search and reranking tasks, which involve ordering results by relevance, it achieved an average result of 83.99 nDCG@10. The metric compares the quality of the top results, and the model exceeded Qwen3-Reranker-8B by 1.81 points.

The model also outperformed Jev 1.14 in all eight languages in a multilingual test: German, Spanish, French, Italian, Russian, Japanese, Korean, and Chinese. Its average accuracy per task was 88.16%, compared with Jev's 85.41%.

In entity extraction, personal information, and questions with extractable answers, Decider v1.3 achieved an average score of 69.0, above the specialized models compared by Perplexity. The Flash version scored 63.6.

In visual tasks, such as classifying invoices, letters, résumés, satellite images, or counting animals, Decider v1.3 reached an average accuracy of 76.99% across five tests. Flash reached 72.77%.

What it can be used for

The practical value lies in automating small, repetitive decisions within larger systems. Perplexity says it uses Decider to:

  • Classify queries and decide whether they need web search.
  • Filter out pages containing spam, CAPTCHAs, errors, or little useful content.
  • Detect phishing, prompt injection, and dangerous behavior.
  • Rank documents by their relevance to a search.
  • Review screenshots for broken layouts or overlapping text.
  • Estimate the difficulty of a question before starting an investigation.
  • Analyze agent steps during reinforcement training.

In an internal security test, the model analyzed 1 billion tokens from 7,000 agent sessions in one day. Perplexity says the analysis detected failures that led to 14 changes to its rules, six of which were approved and published over the next two days.

The company also evaluated one million real API requests, with users' explicit permission for research. Those requests contained 2,318,887 questions. Decider v1.3 matched the answers from judges based on advanced models in 93.98% of cases, although this comparison is not equivalent to independent human verification.

What changes for you

If you build an AI application, Decider can handle decisions that do not need a written response. That lets you reserve larger generative models for complex tasks and use a specialized model to route queries, moderate content, or extract data from thousands of documents.

Perplexity recommends Decider v1.3 for evaluations with detailed criteria and costly errors. Flash is intended for classifying, filtering, and routing large volumes when price and speed are priorities. In both cases, you should calibrate the probabilities with your own examples: a suitable threshold for detecting spam may not work for moderating content or selecting results.

Both models are available with downloadable weights on Hugging Face under the Apache license, as well as through OpenRouter and Perplexity's decisions API. The next point to watch will be how they perform outside the company's tests, especially on multilingual tasks, sensitive-data extraction, and workflows where an incorrect classification triggers automated actions.