AI Wire

Short briefings on models, tools, agents, and the rules that govern them.

English edition

Models

Perplexity Decisions API arrives with open 27B model

Perplexity launched its Decisions API and open-sourced the 27B decider model under Apache 2.0, targeting automated classification over chat output.

Priced at $0.04 per million input tokens with free Apache 2.0 weights, the system targets pipelines that need clean decisions rather than conversation.

Perplexity launched the Perplexity Decisions API in early October 2026 and released open weights for its pplx-decider-v1-27b model on Hugging Face under the Apache 2.0 license. Fine-tuned on top of Qwen3.8-27B, the model has a narrow job: inspect incoming data and output calibrated probabilities instead of generating sentences. In the cloud, access costs $0.04 per million input tokens, and Perplexity charges nothing for output tokens.

How Perplexity Decisions API works

Unlike standard conversational models, pplx-decider-v1-27b functions as a direct numerical classifier. The endpoint receives a state object containing text, JSON payloads, or images, and answers up to 128 questions at the same time in a single API call. Queries run across three structured formats: “noul” (a 0 to 1 probability score for yes-or-no checks), “choice” (probability distribution across predefined options), and “score” (ratings across ordered rubrics of up to ten tiers).

Native image handling allows systems to process screenshots, product photos, or receipts without running a separate optical character recognition (OCR) step first. Images submitted to the API consume roughly 1,000 input tokens per megapixel. The system handles an input context window of up to 262,144 tokens, with an initial rate limit of 10 requests per second per organization.

Cutting conversational overhead in automation

The model addresses a common headache in software engineering. Many teams use conversational chat models to triage incoming customer support tickets, scan transactions for fraud risk, or route sales leads. That setup is expensive, wastes tokens generating chatty prose, and breaks backend pipelines whenever a model decides to talk instead of returning raw data.

By focusing purely on decisions, the system calculates only the statistical distribution of approved options. The response comes back as a numerical vector that feeds directly into basic code logic, such as raising an alert when a risk score passes a set threshold.

Challenging Jev and managing self-hosting costs

The launch places Perplexity in direct competition with TypeSafe, which released Jev in September 2026 through a team of former OpenAI researchers. While Jev operates as a proprietary, text-only cloud service, Perplexity chose to compete with a permissive open license and native multimodal input.

Teams that plan to run pplx-decider-v1-27b on their own servers will need heavy hardware. With 26.8 billion parameters in BF16 precision, the model requires between 49 and 52 GiB of video memory (VRAM) across CUDA GPUs just to load the weights into memory.

Self-reported metrics and pricing promises

In its Hugging Face model card, Perplexity stated that pplx-decider-v1-27b scored an average of 85.71% across an internal panel of 11 decision tests, compared to 84.51% reported for Jev and 74.76% for baseline Qwen3.8-27B. On the RAGTruth benchmark, which measures factual grounding against documents, the company claimed an 88.80% score against 77.27% for Jev. These numbers are self-reported by Perplexity and have not yet undergone independent academic review across the full test suite.

Chief executive Aravind Srinivas posted on X that Perplexity plans to reduce API pricing in the near future. This statement is treated as a commercial promise (labeled RUMOR regarding actual rollout), as the company has not published an updated pricing table or set a formal date for lower rates.

Sources

Sources

  1. docs.perplexity.ai
  2. huggingface.co
  3. openrouter.ai
  4. requesty.ai
  5. aiweekly.co
  6. x.com
  7. x.com

Markers

  • Perplexity
  • Decisions API
  • open-source
  • multimodal
  • machine-learning

Corrections and editorial policy