Perplexity's 27B decision model returns typed probabilities for classification, routing and rubric grading instead of generated text.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Joint Questions, One Pass: Where Decider V1.2 Earns Its Place
Benchmarks
ClaimBench: Near-Perfect Claim Checking Against 10-K Excerpts
Output Speed
Intelligence Index
MMLU *
GPQA *
HLE *
LiveCodeBench *
Technical Specifications
What the model supports
Comparison
Decider V1.2 or Decider Flash V1.2 for Production Decisions?
Both models use the same Decisions API, request format and text and image support, and output tokens are free on both. Flash is a 4B model distilled from the 27B Decider models. Decider V1.2 is the 27B default. In Perplexity's October 19, 2026 test, a request of about 248,000 input tokens took 53 seconds on Decider V1.2 and 16 seconds on Flash. The decision depends on how nuanced each judgment is and how many questions you ask per input. Perplexity advises testing both on labeled data.
Choose Decider V1.2 when judgments are nuanced, such as detailed rubric grading or checking claims against long source text. Also choose it when a wrong answer is costly, or when you ask three or more questions about one state, because it bills the state once. Choose Decider Flash V1.2 for high-volume decisions with clear criteria like routing, labeling and filtering. It also fits single-question requests and latency-sensitive paths. Its smaller footprint is easier to self-host. Keep Flash only if it meets your accuracy bar.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Multi-field ticket triage
Decider V1.2 answers several questions about one message in a single pass, and the joint format costs about as much as one question. A support message can return a team assignment, a blocked yes/no and a frustration score together. In MindStudio's example, the team call came back at 99.8% confidence. This fits helpdesk routing and workflow dispatch where code acts on thresholds. Review low-confidence, idiomatic or sarcastic messages, where hands-on tests showed lopsided negative scores.
Claim verification against source text
On Vals AI's ClaimBench, Decider V1.2 scored 99.69% at labeling claims as supported, contradicted or unsupported by 10-K excerpts. It reached 98.5% on subtle contradictions. Perplexity names checking a claim against long source text as a use case for this larger model. This fits compliance review, filing QA and grounding checks on generated reports. Instruct it to judge only from the supplied excerpt, and validate on your own document types.
Confidence-gated LLM cascades
Each answer carries a probability, so you can accept confident decisions and escalate the rest. Vals AI's cascade handled 99.7% of claims with Decider V1.2 alone and sent the remainder to Claude Haiku 5.5. Combined accuracy was 99.85%. This fits high-volume review pipelines where a language model is only needed for ambiguous cases. You need a held-out labeled set to tune the threshold, and you should revalidate it whenever the model version changes.
Rubric grading and moderation
Score questions return a probability for each level of an ordered rubric, plus an expected score. Choice questions support up to 255 options, and images can be part of the state. Perplexity recommends the larger model for detailed rubric grading. This fits evaluating LLM outputs and moderation queues. The model gives no explanation for a grade, and vision outputs need calibration on your own images.