Perplexity Decider v1.3 adds locate extraction, keeps $0.02 per 1M tokens

1 day ago

9 min read

Perplexity has released pplx-decider-v1.3 and pplx-decider-flash-v1.3, the latest versions of the open-weight decision models behind its Decisions API. The company announced the release on its official API Platform Forum on November 12, 2026, and the Decisions API changelog lists the two models, a new question type called locate, and an alias switch that moves default traffic to the new versions. The models are available through the Decisions API, on Hugging Face and on OpenRouter.

The headline addition is locate, a span-extraction mode. A locate question describes what to find in a text, such as error codes, order numbers or clauses, and the model returns the matching passages as exact substrings of the input with character offsets and a confidence score. Until this release, the Decider models could only answer three kinds of questions: yes/no probabilities, multiple choice and ordered scores. Locate turns the same model into an extraction tool without moving to a generative LLM.

Prices are unchanged. According to Perplexity's API pricing page, checked on November 14, 2026, the 27-billion-parameter pplx-decider-v1.3 costs $0.02 per 1M input tokens and the 4-billion-parameter pplx-decider-flash-v1.3 costs $0.01 per 1M input tokens, with output tokens free and no per-request fee. The practical cost change sits elsewhere: the flash model now answers several questions about the same input in one pass, so a multi-question request on flash bills the shared input once instead of once per question.

··········

PERPLEXITY DECIDER V1.3 MODELS, SIZES AND LICENSE.

Decider is a family of classification and scoring models. A request sends a state, which is the content under review (a support ticket, a product review, a document, an image), together with a set of named questions, and the model returns probabilities instead of free-form text. The format suits high-volume decisions where a full chat model would be slower and far more expensive: routing, moderation, triage, quality checks and rubric grading.

The Decisions API reference lists the 27B model and the 4B flash model as the two v1.3 tiers. The Hugging Face model card for pplx-decider-v1.3 describes it as the successor of pplx-decider-v1.2, built on a Qwen base model, released under the Apache 2.0 license and stored as BF16 safetensors. The card links the base model as Qwen3.8-27B, while the repository header lists 26B parameters, so the exact count depends on which figure is read; the API documentation uses 27B. The card also states that access to the repository requires a Hugging Face token with permission, so the weights are open under the license but not an anonymous download.

........

Model	Parameters (API docs)	Input price per 1M tokens	Output	Locate support
pplx-decider-v1.3	27B	$0.02	Free	Yes
pplx-decider-flash-v1.3	4B	$0.01	Free	Yes
pplx-decider-v1.2	27B	$0.02	Free	No
pplx-decider-flash-v1.2	4B	$0.01	Free	No

........

Open weights and open source are different things here. The Apache 2.0 license covers the published checkpoint and the inference code that ships with it, while the model card does not describe the training data. Perplexity places its training details, evaluation results and real-world cost figures in a separate blog post linked from the announcement. The model card and the API documentation carry no benchmark table, so the accuracy gain over v1.2 is not quantified on the reference pages developers integrate against.

··········

HOW THE LOCATE QUESTION TYPE WORKS IN THE DECISIONS API.

Locate questions use the same endpoint as every other Decider question, POST https://api.perplexity.ai/v1/decisions, and the same request body. The question needs only a type of locate and a non-empty instructions string describing the target. The response returns a spans array, and each span carries a start offset, an exclusive end offset, the text itself and a score where higher means more certainty. Offsets count Unicode characters, and Perplexity guarantees that each returned text is a verbatim slice of the input, so the condition state[start:end] == text always holds. An empty array means nothing matched.

The documented constraints are specific. Locate works only on the two v1.3 models, and calling it on v1.2 returns a 400 error. The state must be a plain string with no images, a request can contain at most 128 locate questions, and locate cannot be combined with web grounding. The API reference says the extraction head was trained on inputs up to 8,000 tokens and evaluated up to 32,000 tokens; longer inputs are accepted, but Perplexity states that span quality beyond that range has not been measured.

Mixed requests are allowed, so one call can ask whether a ticket is urgent, which department should own it, how severe the problem is and where the order number appears. The model card notes that locate questions run in separate forward passes from decision questions, which adds latency to mixed requests on self-hosted deployments. The illustrative request body below uses only field names documented in the API reference.

{
  "model": "pplx-decider-v1.3",
  "state": "Order 88-4412 arrived with a cracked screen. Error E-204 appears on startup.",
  "questions": {
    "urgent": {"type": "noul", "instructions": "Does the customer need a replacement?"},
    "error_codes": {"type": "locate", "instructions": "Device error codes"}
  }
}

The other three question types keep their v1.2 shapes. A noul question returns the probability of yes, a choice question picks among 1 to 255 named options and returns per-option probabilities with a confidence value, and a score question grades the state on an ordered rubric of up to 10 levels in the API, returning a probability-weighted score. The model card warns that prompts written for v1.1 or v1.2 are not compatible with the self-hosted v1.3 checkpoint, which expects the new prompt format produced by its own pplx_decider package.

··········

DECIDER V1.3 API PRICING, ALIASES AND LIMITS.

Perplexity left the per-token prices untouched, and the comparison below follows the official pricing page and API reference as they read on November 14, 2026. Prices are in USD per 1M tokens; the Decisions API has no cached-input rate and no batch discount.

........

Item	Before (v1.2)	After (v1.3)	Effective
27B input price	$0.02	$0.02	November 2026
4B flash input price	$0.01	$0.01	November 2026
Output tokens	Free	Free	Unchanged
Locate questions	Not available	Available on both models	November 2026
Flash multi-question billing	State billed once per question	State billed once per group	November 2026
pplx-decider-latest	v1.2	v1.3	November 2026
pplx-decider-flash-latest	flash-v1.2	flash-v1.3	November 2026

........

The alias change is the part that needs action. Any application calling pplx-decider-latest or pplx-decider-flash-latest is now served by v1.3 without a code change, and the changelog warns that probabilities can change with the model, so thresholds tuned on v1.2 should be revalidated. Teams that want to stay on the previous behavior must name pplx-decider-v1.2 or pplx-decider-flash-v1.2 explicitly. Requests that still name the older pplx-decider-v1.1-27b or pplx-decider-v1-27b identifiers are routed to v1.2 and billed at $0.02.

The API reference sets the same input limits for both generations: the state plus any single question must fit under 256,000 input tokens, and the request body must stay under 32 MiB. The Decisions API quickstart, which still describes the v1.2 models, lists a rate limit of 10 requests per second per organization shared across models, and Perplexity has not published separate rate limits for v1.3. Web grounding, added to the Decisions API this month, runs one Fast Search per question at $0.001 per search, up to 100 searches per request, and charges $0.001 for each successful page fetch when the state is a URL; search evidence and fetched text are counted as input tokens. Perplexity's pricing page does not list a separate rate for locate questions.

··········

[[ADPLUS_300x250_1]]

WHAT A SUPPORT-TICKET TRIAGE PIPELINE COSTS ON DECIDER V1.3.

The cost difference between versions shows up only when several questions share one input. Take a support desk that classifies 100,000 tickets a day. Each ticket is about 600 tokens, and the pipeline asks three decision questions of about 60 tokens each: whether the ticket reports a defect, which of several queues should own it, and how urgent it is on a four-level rubric. On a model that answers all questions in one pass, each request bills 600 plus 180 tokens, or 780 tokens. On flash-v1.2, which re-counts the state for every question, the same request bills three times 660 tokens, or 1,980 tokens. The figures below were computed from those assumptions and the official prices.

........

Model	Billed tokens per ticket	Tokens per day	Cost per day	Cost per 30 days
pplx-decider-flash-v1.2 (before)	1,980	198M	$1.98	$59.40
pplx-decider-flash-v1.3 (after)	780	78M	$0.78	$23.40
pplx-decider-v1.3 (27B)	780	78M	$1.56	$46.80
Gemini Decision 1	780	78M	$7.80	$234.00

........

For a flash user with three questions per ticket, the switch to v1.3 cuts the monthly bill from $59.40 to $23.40, a 61% reduction, with no change in the listed price. The 27B model costs twice as much as flash-v1.3 per token and, at three questions, now costs more than flash rather than less; under v1.2, Perplexity's own documentation advised the 27B model for three or more questions on one state because flash re-billed the input each time. Teams that had moved to the 27B model only for that billing reason can test whether flash-v1.3 meets their accuracy needs.

Web grounding changes the economics because its per-search fee outweighs the token cost by a wide margin. A fact-checking job that grounds 10,000 claims a day, with one question per claim and an assumed 1,500 tokens of search evidence added to a 400-token claim and an 80-token question, spends about $0.40 a day on tokens with pplx-decider-v1.3 and $10 a day on searches. For grounded workloads, the number of questions per request drives the bill far more than the choice between 27B and flash.

··········

DECIDER V1.3 COMPARED WITH GOOGLE'S GEMINI DECISION 1.

Google entered the same category on November 11, 2026, one day before Perplexity's release, with the Gemini Decisions API and its gemini-decision-1 model, which Data Studios covered in its report on the Gemini Decisions API launch. According to Google's announcement, Gemini Decision 1 costs $0.10 per 1M input tokens with output free, supports noul, choice and score questions, accepts up to 256,000 tokens of text, images, audio, video and PDFs, and allows up to 1,000 questions per request with the shared input billed once.

........

Feature	Perplexity Decider v1.3	Gemini Decision 1
Input price per 1M tokens	$0.02 (27B), $0.01 (flash)	$0.10
Output tokens	Free	Free
Question types	noul, choice, score, locate	noul, choice, score
Questions per request	Up to 128 (100 with web grounding)	Up to 1,000
Inputs	Text and images (locate: text only)	Text, images, audio, video, PDFs
Open weights	Yes, Apache 2.0	Not stated
Web grounding	$0.001 per search	Not described in launch post

........

The comparison is not like-for-like on quality, because neither company has published results on a shared benchmark that both models were run against. On published list prices, Decider v1.3 costs one fifth of Gemini Decision 1 per input token on the 27B tier and one tenth on flash, and it is the only one of the two with span extraction and downloadable weights. Gemini Decision 1 accepts audio, video and PDF inputs directly and allows far more questions per call, and Google reports end-to-end latency of 150 to 250 milliseconds, a figure Perplexity has not published for the hosted API.

··········

WHAT CHANGES FOR DEVELOPERS USING THE PERPLEXITY DECISIONS API.

Developers already on the Decisions API should start with the alias. A team that relies on pplx-decider-latest received v1.3 automatically, and any fixed cut-off, such as routing a ticket to a human above a 0.8 probability, was tuned on a different model. A practical way to follow the changelog's advice is to replay a labeled sample through v1.3, compare probability distributions with the v1.2 results, and either retune thresholds or pin the v1.2 model name until the check is done.

Flash users with more than one question per request get the largest saving and should verify their bills after the switch. Pipelines that combined Decider with a separate extraction step, such as a regular expression or a generative model pulling out IDs and dates, can test whether a locate question in the same request replaces that step, keeping inputs under the 8,000-token range on which Perplexity trained the extraction head. Self-hosting is possible under Apache 2.0, but the model card asks for a CUDA GPU with room for about 49 GiB of BF16 weights plus working memory and Python 3.12 or later, which puts the 27B checkpoint out of reach of most single consumer GPUs.

Perplexity's wider API, including Agent API effort levels and Search API pricing, is described in our guide to the Perplexity API and its pricing. The Agent API also changed this month, as covered in our report on the ultra effort level and the medium-effort price cut.

··········

PERPLEXITY DECIDER V1.3: CONFIRMED FACTS AND OPEN QUESTIONS.

The documented facts are clear. Perplexity shipped pplx-decider-v1.3 and pplx-decider-flash-v1.3 on November 12, 2026, kept prices at $0.02 and $0.01 per 1M input tokens with free output, added a locate question type that returns verbatim spans with offsets, moved the default aliases to v1.3, and changed flash billing so that several questions on one state count that state once. The 27B checkpoint is published on Hugging Face under Apache 2.0.

Several points remain unverified from the official pages that could be read. The accuracy of v1.3 against v1.2 and against other models is reported only in Perplexity's blog post and not in the model card or API reference, the parameter count is listed as 26B in one place and 27B in others, the pricing page does not say how locate questions are billed when mixed with decision questions, and the hosted API's latency is not published. Against Gemini Decision 1, v1.3 is the cheaper option on list price; whether it is the better one for a given pipeline depends on an accuracy test against that team's own labeled data, the same check Perplexity's changelog asks for before relying on new thresholds.

··········

·····

DATA STUDIOS

·····

datastudios.org