Perplexity Decider v1.3: Locate, API Pricing, Open Weights

Last updated: November 13, 2026

Perplexity Decider v1.3 and Decider Flash v1.3, released November 12, 2026, are decision models that return probabilities instead of text. The new version adds locate, which extracts exact text spans with character offsets. Both are in the Decisions API at $0.02 and $0.01 per million input tokens, with free output tokens and Apache 2.0 weights.

Contents

What is new in Decider v1.3
Model names, prices and limits
How to call Decider v1.3
Using locate for span extraction
Running the open weights
Decider v1.3 vs. Flash v1.3 vs. v1.2
Limitations
Frequently asked questions

What is new in Decider v1.3

Perplexity announced the release in its community forum on November 12, 2026. The API changelog lists three changes:

New models: pplx-decider-v1.3 (27B) and pplx-decider-flash-v1.3 (4B), at the same prices as v1.2.
New question type, locate: returns the parts of a text input that match a description, with character offsets and a score for each span.
Alias changes: pplx-decider-latest now selects v1.3 and pplx-decider-flash-latest selects Flash v1.3.

Both v1.3 models also answer all questions in a request in one pass, with each question isolated from the others, so the input (called state) is billed once. If you are new to decision models, start with our Perplexity Decisions API guide, which covers v1.2.

Model names, prices and limits

The figures below come from the Decisions API documentation.

Item	pplx-decider-v1.3	pplx-decider-flash-v1.3
Alias	pplx-decider-latest	pplx-decider-flash-latest
Parameters	27B	4B
Input price	$0.02 per 1M tokens	$0.01 per 1M tokens
Output tokens	Free	Free
Input types	Text and images	Text and images
Context window	256K tokens	256K tokens
Options per choice / levels per score	255 / 10	255 / 10
locate questions	Yes	Yes

Other documented limits: a request body of up to 32 MiB, up to 8 images per question, and up to 100 questions when web grounding is on. Web grounding costs $0.001 per search or successful URL fetch, plus input tokens for the added evidence.

How to call Decider v1.3

The endpoint is a single JSON POST to https://api.perplexity.ai/v1/decisions. Steps from the quickstart:

Create an API key at console.perplexity.ai and set it as PERPLEXITY_API_KEY.
Install the Python library (perplexityai 0.44.0 or later) or the TypeScript SDK (0.39.0 or later). Any HTTP client also works.
Set model to pplx-decider-v1.3, or to pplx-decider-flash-v1.3 for Flash. Use the -latest aliases only if you want automatic upgrades.
Pass your content as state (a string, object or array) and add named questions of type noul (yes/no), choice, score or locate.
Read answers by question name and compare each probability with a threshold you tune on your own labeled data.

A noul answer is the probability of yes. A choice answer returns the top option, a probability for every option and a confidence value. A score answer returns the expected level on your rubric.

Using locate for span extraction

A locate question takes a short label such as “email address” or an instruction such as “Extract the city where the job is located.” The answer is a list of spans, each with text, start, end and score. The text is a verbatim substring of the input, so state[start:end] equals the span in Python. No match returns an empty list.

Rules from the documentation:

Only the two v1.3 models answer locate; v1.2 returns HTTP 400.
state must be a plain string. Images or object and array inputs return 400.
A request with a locate question can hold up to 128 questions and can mix in noul, choice and score questions.
locate does not work with web grounding.
Locate was trained on inputs up to 8K tokens and evaluated up to 32K tokens. Longer inputs are accepted, but span quality on them is not measured.

Typical uses are named entities, personal information detection and extractive question answering, for example pulling an error code out of a support ticket while a noul question decides whether the customer is blocked.

Running the open weights

Both models are on Hugging Face under the Apache 2.0 license. According to the Decider v1.3 model card and the Flash v1.3 model card:

Decider v1.3 is built on Qwen3.8-27B and needs a CUDA GPU with room for about 49 GiB of BF16 weights plus working memory.
Decider Flash v1.3 is built on Qwen3.5-4B and needs room for about 8.5 GiB of BF16 weights.
Both need Python 3.12 or later and the bundled pplx_decider package. The cards say the models expect the v1.3 request format, so do not reuse v1.1 or v1.2 prompt builders.
Requests that mix locate and decision questions need separate forward passes, which adds latency.
The local default max_length is 32,768 tokens, and training used sequences up to 30,720 tokens.

The models are also on OpenRouter; its Decider V1.3 (free) page lists a 262K context window and notes that free endpoints are rate limited.

Decider v1.3 vs. Flash v1.3 vs. v1.2

Feature	Decider v1.3	Decider Flash v1.3	Decider v1.2 / Flash v1.2
Size	27B	4B	27B / 4B
Input price per 1M tokens	$0.02	$0.01	$0.02 / $0.01
locate questions	Yes	Yes	No (returns 400)
Question grouping	One pass, questions isolated	One pass, questions isolated	v1.2 groups; Flash v1.2 answers each question on its own
Recommended use	Default; nuanced rubrics, claim checks, costly errors	High-volume routing, labeling, filtering, simple extraction	Only while you revalidate thresholds

Perplexity’s advice is to run both v1.3 models on a sample of your own labeled data and keep Flash only if it meets your accuracy bar. For a competing approach, see our Gemini Decision 1 guide.

Power-user workflows

Ticket triage: combine a choice question for the team, a noul question for urgency and a locate question for order numbers in one request.
Agent gating: score proposed actions with Decider before a generative agent runs them, then send only hard cases to the Perplexity Agent API.
Fact checks: use web grounding with noul questions, covered in our Agent Ultra and web grounding guide.
Local testing: compare hosted and local results, much like the decision-model support described in our Ollama features guide.

Limitations

Decider does not write replies, generate code or explain its reasoning; your code must act on the numbers.
Probabilities can change between versions, so thresholds tuned on v1.2 need retesting.
locate is text-only and cannot be combined with web grounding.
The documentation does not publish a retirement date for v1.2.

Frequently asked questions

What is Perplexity Decider v1.3?

Decider v1.3 is a Perplexity decision model released November 12, 2026. Instead of writing text, it returns probabilities for yes/no, choice and score questions about text, JSON or images, and it adds locate, which returns matching text spans with character offsets.

How much does Decider v1.3 cost on the Decisions API?

pplx-decider-v1.3 costs $0.02 and pplx-decider-flash-v1.3 costs $0.01 per million input tokens, the same as v1.2. Output tokens are free. Web grounding adds $0.001 per search or URL fetch.

What is the difference between Decider v1.3 and Decider Flash v1.3?

Decider v1.3 is the 27B default model for nuanced judgments and costly errors. Decider Flash v1.3 is a 4B model for high-volume routing, labeling, filtering and simple extraction at half the input price. Both have a 256K context window and support locate.

Can I download the Decider v1.3 weights?

Yes. Both models are listed on Hugging Face under the Apache 2.0 license. The model card asks for Python 3.12 or later and a CUDA GPU with room for about 49 GiB (v1.3) or 8.5 GiB (Flash v1.3) of BF16 weights plus working memory.

Do I need to change my code to move from v1.2 to v1.3?

Requests that use pplx-decider-latest or pplx-decider-flash-latest now go to v1.3 automatically. The request format is unchanged, but probabilities can shift, so Perplexity says to revalidate thresholds. Pin pplx-decider-v1.2 while you test.

Perplexity Decider v1.3: Power-User Guide to Locate, Pricing and Weights