Perplexity pplx-decider-v1.2 adds 4B Flash decision model

What was released?

Perplexity's Decisions API now offers two new decision models: the 27B-parameter pplx-decider-v1.2 and the new 4B-parameter pplx-decider-flash-v1.2 (Perplexity changelog). v1.2 replaces pplx-decider-v1.1-27b. Vercel AI Gateway and OpenRouter added both models on October 16 (Vercel, OpenRouter).

A decision model does not write text. It takes text, JSON or images plus one or more questions, and returns a probability for every possible answer (Perplexity). There are three question types:

• yes/no: the probability of yes, from 0 to 1;
• choice: one of the options you define, with a probability for each;
• score: a level on an ordered rubric, with an expected score.

Perplexity pitches it for cases where you would otherwise ask a chat model for a label and parse the reply: classification, routing and grading against a rubric.

Pricing and limits

According to the official pricing page, pplx-decider-v1.2 costs $0.02 and the Flash model $0.01 per million input tokens (Perplexity pricing). Output tokens are free and there is no per-request fee. A 600-token request costs $0.000012 on the large model and $0.000006 on Flash.

Both models take the same request, so switching is a one-line change to the model field. Shared limits:

• 256K-token context window;
• up to 128 questions per request;
• up to 255 options per choice question and 10 levels per score question;
• up to 8 images per question;
• 10 requests per second per organization, on every plan.

The pplx-decider-latest and pplx-decider-flash-latest aliases follow the newest version. The old v1.1-27b and v1-27b model names still work. They now select v1.2 and bill at its rate.

What changed from v1.1?

According to the Hugging Face model card, v1.2 uses the same Qwen3.8-27B backbone (Hugging Face). Two things changed: the request format and the training data. v1.1 answered one question per forward pass. v1.2 reads the state and all questions in one sequence. Several questions now cost about as much as one, and the questions can inform each other.

Decision Index 0.3 results reported by Perplexity:

• overall Decision Index: 68.09, up from 62.30 for v1.1 (67.75 with one question per request);
• Knowledge: 56.3, up from 51.8;
• Retrieval: 68.4, up from 61.2;
• Arts: 56.2, up from 47.3;
• Vision Decision Index: 77.43, up from 73.22.

Grouping questions is not always a win. When the questions are facets of one decision, it helps: +11.3 points on SATA-Bench. When they are competing candidates, it hurts: −4.8 points on ToolRet. Perplexity therefore suggests one candidate per request when in doubt.

On calibration, the model card reports an expected calibration error of 0.014 on the development set. On decision data held out from all training, probabilities remain somewhat overconfident (about 0.05), though less so than v1.1.

Open weights?

Perplexity announced v1.1-27b as an open-weight model (Perplexity Community). The v1.2 Hugging Face repository, however, is private according to its model card, and downloading requires a token with access. The sources do not confirm a public, free download for v1.2. We found no Hugging Face model card for the Flash model.

Availability: what the sources confirm

• Perplexity API: available, callable with any Perplexity API key at POST /v1/decisions.
• Vercel AI Gateway and OpenRouter: both models are available.
• Pricing: listed on Perplexity's official pricing page, see above.
• Perplexity consumer app (search, Pro): the sources say nothing about the model appearing there. This is not confirmed.

What it means in practice

The decision model market is moving fast. Liquid AI's d1-ultra, released a day earlier, costs $0.042 per million input tokens (Liquid AI). Liquid's comparison table still lists only v1.1 of Perplexity's model, and there is no independent measurement of v1.2 yet.

For high-volume, simple routing or labeling, Flash is the cheapest entry point. Perplexity itself recommends testing both models on your own labeled sample first and keeping Flash only if it meets your accuracy bar.