Perplexity Decider v1.3: what's new and what it costs

Perplexity Decider v1.3 is the new version of the open-weight model behind the Perplexity Decisions API, released November 12 alongside a smaller Decider Flash v1.3. The main addition is locate, which pulls exact pieces of text out of a document. Through the API it costs $0.02 per million input tokens, Flash costs $0.01, and output is free.

If you're new to this, start with our first post on the Perplexity Decisions API. The short version: you send some text or images plus a list of fixed questions, and the model returns probabilities instead of prose.

The X posts behind it

Two posts from October set this release up. CEO @AravSrinivas announced that Decisions API pricing "was cut by half, now to 2 cents per million input tokens." Then @perplexitydevs shipped v1.2 in two sizes and said inference was "up to 5x faster" because every question in a request runs in one pass.

I couldn't find a matching X post for v1.3 today. The announcement I can verify is on Perplexity's developer forum, which says v1.3 "adds locate, a new feature that extracts spans based on type or instructions," with the details in a longer blog post.

What locate does

The older question types all pick from answers you define: yes or no, one of several options, or a score on a rubric. With locate you describe what you want instead, and the model returns the matching text copied word for word from the input.

The quickstart shows a locate question as one small object:

{"type": "locate", "instructions": "Every product name mentioned in the review."}

Because the answer has to be a span from the source, the model can't paraphrase a dollar amount or invent a name that isn't there, which is what you want when the result is headed for a spreadsheet or a CRM field.

It works on text only, though. A request can carry at most 128 locate questions, and you can't combine locate with the API's web grounding.

What it costs

The prices are from Perplexity's docs and the memory figures are from the Hugging Face model cards for v1.3 and Flash. Both are tagged Apache 2.0. The big model is built on Qwen3.8-27B, and Flash is distilled from it into a 4B model. Web grounding, if you turn it on, adds $1 per 1,000 searches or page fetches.

At these prices, 10,000 emails of about 500 tokens each cost roughly 10 cents on v1.3 and 5 cents on Flash.

How to try it

1. Get a Perplexity API key and install the perplexityai Python package, version 0.44.0 or later.
2. Call the decisions endpoint with the model set to pplx-decider-latest, which now points at v1.3.
3. Send one real document as the state, such as a customer email or a vendor invoice pasted as text.
4. Add a locate question. Some I'd start with: "The invoice number," "Every date the customer says they are available," "The sentence where the customer asks for a refund."
5. Add a yes/no question in the same request, such as whether the message is urgent, and compare both answers against what a person on your team would say.

Perplexity suggests the larger model "for nuanced judgments, detailed rubrics, and workflows where errors are costly." If you mostly care about volume, I'd test Flash first.

How good is it?

The numbers are Perplexity's own. On 17 locate tasks outside its training domains, the blog reports an average score of 69.0 for v1.3 and 63.6 for Flash, against 43.3 for Liquid's d1 and 41.5 for Cloudflare Clairic. On a million real API requests, agreement with an LLM judge was 93.98% for v1.3, 93.68% for v1.2 and 93.35% for Jev 1.14.

The locate lead is wide, but the gain in decision quality over v1.2 is 0.3 points, so if you only use yes/no and choice questions there's no rush to upgrade. Perplexity also admits most of these tests come from its internal suite and that leaderboards "can still be gamed." I wouldn't trust any of these figures until I'd run the model on my own documents.

What this means for a small business

Sorting messages was already cheap. Locate goes after the step that comes next, which is retyping: someone copies the invoice number or the requested date out of an email and into another system. That work is dull and easy to get wrong. It now costs about a dime per ten thousand documents, on a model you could also run on your own hardware.

If you want to know which of those copy-and-paste steps are worth automating first, New Face Design's free process audit maps where your team's hours go.