Perplexity’s Decider v1.2 returns probabilities for classification and scoring rather than generated prose. Its hosted Decisions API costs $0.02 per million input tokens; Flash costs $0.01, with output tokens uncharged. The key migration change is joint requests: several questions share one state, and their answers can influence one another.
That changes how developers should package a workload. Asking about a ticket’s department and urgency can be a coherent joint decision. Scoring unrelated search candidates together introduces a different dependency. A larger batch is not automatically an equivalent, cheaper implementation of the original independent judgments.
Choose the question format before the model
The Decisions API documentation defines noul for a yes/no probability, choice for a distribution over named alternatives, and score for an ordered rubric. The endpoint is POST https://api.perplexity.ai/v1/decisions.
For an application, this removes a parsing step: software receives typed numbers rather than trying to recover a label from a chat response. It still needs a decision policy. A classification result and permission to act on that classification are separate application concerns.
A simple routing request, using an existing PERPLEXITY_API_KEY, looks like this:
curl https://api.perplexity.ai/v1/decisions \
-H "Authorization: Bearer $PERPLEXITY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "pplx-decider-v1.2",
"state": "The replacement shipment arrived damaged; please arrange another delivery.",
"questions": {
"department": {
"type": "choice",
"instructions": "Which queue should receive this message?",
"criteria": {
"delivery": "Damaged or missing shipments",
"billing": "Invoices or payment failures",
"other": "Anything else"
}
}
}
}'This is a documented request pattern with an original example, not a reported API test. Save the response’s complete distribution when evaluating it; a winning label alone hides whether the alternatives were nearly tied.
Group related attributes; examine candidate competition
The 27B checkpoint’s model card explains that the joint format places a readout marker after each question. Full-attention layers allow questions to see one another. The card reports improvements from grouping related attributes, while some candidate-retrieval tasks decline.
Its overall Decision Index score rises from v1.1’s 62.30 to 68.09 with joint requests. V1.2 sent one question at a time scores 67.75. The arithmetic difference between those two v1.2 configurations is 0.34 points. Most of the version-to-version improvement should therefore not be attributed simply to combining questions.
For migration, keep an identical held-out set and compare one-question requests against the proposed groups. Record grouping membership and question order. For search reranking, include closely similar distractors, because that is where a small probability shift may change the top result. Compare ranking quality and cost together, rather than requiring identical probabilities from two different request layouts.
The Flash model card supplies a useful counterexample to a universal grouping rule: its overall Decision Index result is 58.88 jointly and 60.05 one question per request. It is a Qwen3.5-4B-based model, distilled from larger decision models. The smaller model’s preferred packaging needs its own evaluation.
Confidence is not the top probability
The API defines choice.confidence as a normalized distance above a uniform distribution, not the probability of the winning option. With four options and a top probability of 0.70, its documented formula gives (0.70 − 0.25) ÷ 0.75 = 0.60.
That distinction matters when replacing an older classifier. A rule written for a 0.70 winning probability cannot be copied unchanged to a 0.70 confidence threshold. Store both fields, decide which quantity the policy actually needs, and recalibrate on labeled examples from the deployment domain.
An illustrative policy might automatically route only high-certainty messages and send the remainder to a review queue. The appropriate cutoff depends on the cost of misrouting and the proportion of traffic the team can review. Evaluate that trade-off using false-route counts and review volume, rather than declaring a model-generated number a guarantee of correctness.
Hosted access and local deployment have different limits
The hosted API is the shortest integration path. At the published input rates, one million requests averaging 1,000 billed input tokens each would contain one billion input tokens: $20 for the larger model or $10 for Flash. This calculation excludes other services and does not establish a production throughput allowance.
Local use has a different resource budget. The larger checkpoint’s documentation calls for Python 3.12+ and approximately 52 GiB of BF16 weights plus working memory. Flash lists approximately 9 GiB plus working memory. Neither figure is a complete GPU-memory guarantee for an arbitrary batch or context length.
The native checkpoints also use a specialized decision readout, rather than an ordinary text-generation head. Use their provided loader and request format; do not assume that a generic chat server reproduces the same probabilities. Although the cards identify Apache 2.0 licensing, their usage instructions mention authenticated repository access, so confirm current download access before allocating hardware.
The useful adoption test is narrow: does the model make your specific structured decision accurately enough, at the right review threshold, with the intended question layout? That result is more informative than either a leaderboard position or a very small per-token price alone.
