Using Perplexity Decider v1.1: review classification, support routing, and cost

Which tasks suit this update?

Perplexity's official announcement account said the updated Decider v1.1 was available on October 6, 2026. The post displays 9:11 PM without a time zone, so it cannot be converted directly to Beijing time or used to assign every region a separate launch date.

It serves stages needing a clear decision: review classification, ticket routing, and scoring against criteria. The model reads input and returns probability-bearing results. A suitable generative model can handle writing replies or explanations afterwards.

The practical benefit is clearer responsibility. Routing a complaint to the correct queue before a support worker handles it separates permissions and roles more cleanly than having one model classify, explain, and execute a refund together.

Choose a decision type before writing the question

noul asks whether a condition holds and returns the affirmative probability. choice evaluates your supplied categories and their probabilities. score uses ordered levels and returns a weighted score. They solve different problems; merely changing a field name is insufficient.

Fictional example: a buyer says, “It looks nice, but the charging port does not work at all.” You could separately ask whether a fault is reported, whether product or logistics support should handle it, and whether severity is cosmetic, inconvenient, or unusable. This is not an actual API result.

Define confusing boundaries first: distinguish late delivery from product failure and decide what to do with insufficient information. Review misclassifications on samples you are authorized to process, then set routing conditions rather than copying someone else's probability threshold.

What do you need for integration?

The hosted endpoint is POST https://api.perplexity.ai/v1/decisions with model pplx-decider-v1.1-27b and Authorization: Bearer authentication. state carries content and questions defines decision questions. Consult the linked official documentation for complete examples.

Perplexity says a valid API key works for this endpoint. Existing teams should check the owning project and organization; project creation, key management, and billing require appropriate permissions. Keep keys server-side or in secure local configuration, never in a public page.

The API uses prepaid credits. Exhausted balance blocks further calls until replenished. Being able to log into Perplexity or download public weights does not provide hosted API balance. Costs belong to the organization owning the key.

What does $0.02 mean for one task?

At review time, the official pricing page lists both v1.1 and v1 at $0.02 per million input tokens, free output, and no per-request fee. Input includes content, questions, and images. Check usage.input_tokens in the response for actual billed quantity.

A hypothetical request with 600 total input tokens costs $0.000012 at that rate. This is arithmetic, not an API call performed here. Text length, question count, and images affect real cost, so review counts alone cannot determine a budget.

The October 6 announcement described a halving using v1's old price, but the current page lists both versions at the same rate. Do not continue promoting v1.1 as half the price of v1; use current pricing and actual usage when choosing and budgeting.

How do you avoid incorrect routing after integration?

Each organization currently has a limit of 10 requests per second and an input-token burst limit. For 429 responses, honor Retry-After and bound retries. Another key in the same organization should not be treated as extra concurrency allowance.

Images accept PNG, JPEG, and WebP base64 data URLs with size limits; ordinary HTTPS image links return 400. Prepare compliant images before calling. An upload failure or timeout is not a model judgment that nothing is wrong.

Inspect HTTP status before error.message and error.type; do not rely solely on error.code, whose type varies. For 401, check Bearer and the key; for 404, check path and trailing slash. Failed requests or uncertain judgments can enter a human queue. That is your business fallback, not an automatic API refusal feature.

confidence for choice and score is the model's own certainty, not measured accuracy in your business. Keep an “other” or review path, especially for refunds, permissions, and important account actions. Successful classification supplies a result; it does not authorize downstream operations.

Do open weights mean free local execution?

Perplexity also offers open weights. Current Hugging Face metadata marks perplexity-ai/pplx-decider-v1.1-27b public and non-gated under Apache 2.0. Its model card was anonymously readable; an older private notice should not establish that weights remain closed.

The card's inference example requires Python 3.12 or newer and CUDA. Weights occupy approximately 49 GiB, with additional runtime space required. Specialized decision-output and calibration processing is involved; downloading files does not establish compatibility with an ordinary chat-serving stack.

The hosted API avoids maintaining inference infrastructure. Self-hosting carries hardware, configuration, and maintenance costs. Assess data-processing requirements and resources before choosing; a public license is not a zero-cost promise, and no local deployment or performance test was performed here.