Perplexity's Decisions API Answers With Probabilities Instead of Text, at $0.04 per Million Input Tokens. What It Does and Where It Stops
Send content plus up to 128 named questions and get back yes/no probabilities, choice distributions, or scored rubrics. Output tokens are free, and only one model is available.
TL;DR
Perplexity's Decisions API, documented for the model pplx-decider-v1-27b, returns probabilities rather than generated text. You send content and 1 to 128 named questions of three types: yes/no, multiple choice, or a scored rubric. Input costs $0.04 per million tokens and output tokens are free. Requests can be up to 262,144 tokens, with a rate limit of 10 requests per second per organisation. Perplexity's documentation does not state a benchmark score; an 85.71% figure circulating in news roundups could not be traced to a method.
Most language model APIs return text, and developers then parse that text to decide what to do. Perplexity's Decisions API skips the parsing: you ask a question about your content and get back a number.
How it works
You POST to https://api.perplexity.ai/v1/decisions with the model pplx-decider-v1-27b, a state (text, a JSON object, or an array, which can include base64-encoded images), and 1 to 128 named questions, each with a type and criteria. The response contains an answer per question name.
| Question type | What you ask | What comes back |
|---|---|---|
| Yes/no | Does this content meet a condition? | A probability from 0 to 1 |
| Choice | Which of your options fits? | The most likely option, probabilities for every option, and a confidence value |
| Score | Where does this land on your rubric? | A probability-weighted average level, the full distribution, and a confidence value |
Price and limits
- Price: "$0.04 per million input tokens. Output tokens are free." Input tokens are reported in the response's usage.input_tokens field.
- Context: under 262,144 input tokens per request, with a 32 MiB maximum body.
- Images: PNG, JPEG, and WebP, sent only as base64 data URLs (no HTTP or HTTPS links). Perplexity documents 2,048 tiles of 32 by 32 pixels, at about 1,000 tokens per megapixel.
- Rate limit: 10 requests per second per organisation.
- Models: only pplx-decider-v1-27b is available.
What it isn't telling you
The quickstart we read does not give accuracy figures or a benchmark. AI Weekly's roundup cites an 85.71% score for the model, but we couldn't find that number in Perplexity's documentation or a description of how it was measured, so we aren't relying on it. Probabilities from a model are not guarantees; the usable test is whether they are calibrated on your content, which you have to check yourself.
What this means if you build with models
Good fits are the jobs where you currently ask a model for yes, no, or a label and then parse the sentence: routing support tickets, flagging policy violations, grading drafts against a rubric. Because you get a probability, you can set your own threshold and send borderline items to a person. At $0.04 per million input tokens with free output, a pilot over a few thousand of your own documents is cheap; run it against a labelled sample before wiring it into anything that acts automatically.
AI Industry Reporter
Priya covers model releases, industry announcements, and the gap between what labs claim and what independent evaluators actually find. She reads the primary source - the paper, the system card, the benchmark org's own statement - before writing a word.
More on AI Automation & Agents
Inworld Buys Ultravox: What Changes for Voice-Agent Builders Today, and the Pricing and Voice Promises Nobody Has Made in Writing
Priya Nair · 5 min
OpenAI and Synopsys Announce GPT-Synopsys, a Model That Runs Chip-Design Software. What's Confirmed, and What Isn't Priced or Dated Yet
Priya Nair · 5 min
OpenAI's Dots Are Always-On Agents With Their Own Cloud Computer. Here's What They Can Do, What They Hand Back to You, and What Isn't Clear Yet
Priya Nair · 6 min