An API that answers with probabilities
Perplexity documents a Decisions API at POST https://api.perplexity.ai/v1/decisions. You send content as state (text, JSON or images) plus named questions. Each answer is a probability, never free text.
Question types are:
- noul, a yes or no probability from 0 to 1
- choice, one option from a map you define
- score, a level on an ordered rubric
The only permitted model id is pplx-decider-v1-27b. The API reference says a missing or unknown model returns 400.
Limits in the docs include 1 to 128 questions per request, under 262,144 input tokens, and 10 requests per second per organization.
Another open decision model after Jev and Clef
Decision models became a product category in September with TypeSafe's Jev. Cloudflare shipped Clef on 1 October. See Cloudflare Clef and Clef-flash: Apache 2.0 multimodal decision models on Workers AI, with a 64k context window.
Perplexity's hosted path sits next to OpenAI's limited Decisions API preview from DevDay. See OpenAI DevDay 2026 brings Ultrafast, a Decisions API preview, Codex cloud and plugin extensions.
The open weights give teams a local option in the same family as Strands Decider and Clef.
Cheap hosted calls and Apache 2.0 weights
The docs price the API at $0.04 per million input tokens. Output tokens are free.
The Hugging Face card carries Apache 2.0 and names Qwen3.8-27B as the base. When The Frontier checked at 19:07 SGT on 5 October, the repo had 81 likes. The card says local runs need about 49 GiB of weights plus working memory.
Pick it for cheap classification at volume
Perplexity reports an 11-benchmark overall of 85.71% for pplx-decider-v1-27b against 84.51% for Jev. The Frontier found no independent test. The card says those scores were measured through the Perplexity API.
| Benchmark | pplx-decider-v1-27b | Jev |
|---|---|---|
| Overall | 85.71% | 84.51% |
| RAGTruth | 88.80% | 77.27% |
| JudgeBench | 78.29% | 78.57% |
| JevBench public hard | 70.30% | 73.27% |
| WinoGrande | 83.30% | 90.70% |
Pick the API when you want typed answers at $0.04 per million input tokens and can stay inside 10 requests per second. Pick the weights when you can host a 27B decision head yourself.
Images must be base64 data URLs. Remote http image URLs return 400.
Rows Jev still wins, and a one-minute image trap
Jev leads several rows in Perplexity's own table, including WinoGrande, JudgeBench and JevBench public hard.
Oversized images do not fail fast. The docs say a request over the 2,048 tile limit waits about a minute and then returns 504. Resize before you send.
Identical requests usually match, the docs say, but can differ in the second decimal place. Set thresholds with margin.
Docs, API reference and weights
The quickstart, the POST /decisions reference and the Hugging Face card are listed under Sources below.