# Perplexity Decider v1.2 and Flash: Pricing, Grouping and Threshold Changes Perplexity's updated decision models add joint question processing and a cheaper Flash option, with practical consequences for existing API integrations. By Paras Madan | 2026-10-18 Perplexity [announced Decider v1.2 and Decider Flash v1.2](https://x.com/perplexitydevs/status/2111166029539758120) on October 16 (UTC). The release updates its 27B decision model and adds a smaller 4B option for applications that classify content or choose between predefined outcomes. For existing users, the important detail is that older API model names now select the new version. These models return probabilities for yes-or-no questions, choices and scoring rubrics. That makes them useful for something like a support queue, where software needs to decide which team should receive a ticket and whether it needs review. The application still owns the rules for acting on those numbers. ## Decider v1.2 and Flash pricing The [official pricing page](https://docs.perplexity.ai/docs/getting-started/pricing) lists Decider v1.2 at $0.02 per million input tokens and Flash at $0.01. Both have free output tokens and no per-request fee. Content, questions and images all contribute to input usage. For a hypothetical million requests averaging 1,000 billed input tokens each, that works out to $20 for the larger model or $10 for Flash. Those are token charges, before the rest of the application's costs. At that volume and request size, even a small increase in mistakes requiring human review could matter more than the $10 saving. The useful comparison is therefore the cost of correctly handled work. A team should count rerouted tickets and manual corrections alongside its API bill, especially if Flash is being considered for a workflow where errors create extra work elsewhere. ## Grouped questions can change the answer The new models can answer several questions about the same content in one pass. Each question can see the other questions, and the shared content is read and billed once. This can help when the questions describe related aspects of one item. Perplexity warns that grouping independent competing candidates, such as separate documents being judged for relevance, can reduce quality. Consider a support ticket describing a failed payment. Asking about its category and urgency together could be a sensible experiment. Scoring ten possible knowledge-base documents in one request is a different experiment: adding a new candidate could change how the existing candidates score. A retrieval pipeline needs to test that behavior before treating grouped scores as interchangeable with independently calculated scores. Flash deserves separate attention. Perplexity recommends one question per request as its default, with grouping reserved for closely connected decisions that have been tested on the actual task. Copying the larger model's request layout into Flash can therefore change more than the price. ## Existing integrations need a threshold check The [Decisions API documentation](https://docs.perplexity.ai/docs/decisions/quickstart) says requests using `pplx-decider-v1-27b` or `pplx-decider-v1.1-27b` now select `pplx-decider-v1.2`. The response's `model` field echoes the requested name, so an old name in a log does not establish that the old model served the request. The `latest` aliases also follow future model updates. For a system that sends a ticket to a human when a score crosses a cutoff, that deserves a regression test even if the API call still succeeds. Save the content, question wording and grouping used in evaluation. Otherwise, a changed result could come from the new model, a revised question or a different set of neighboring questions. Start with labeled examples that include ambiguous cases and the errors the team most wants to avoid. Compare both models at the same acceptable error rate, then measure how much work each can handle automatically. Inspect the cases close to the cutoff separately. A higher average accuracy can coexist with a worse result in the narrow score range that triggers an action. The threshold is a product decision too. Routing something to the wrong internal queue and automatically closing a valid support request have different consequences. They should have different evaluation criteria, even if both use the same model. ## Open weights and context limits Perplexity provides Apache 2.0 weights for the [27B Decider model](https://huggingface.co/perplexity-ai/pplx-decider-v1.2), built on Qwen3.8-27B, and [Decider Flash](https://huggingface.co/perplexity-ai/pplx-decider-flash-v1.2), built on Qwen3.5-4B. Their model cards specify approximately 52 GiB and 9 GiB of BF16 weights respectively, plus working memory. Self-hosting therefore needs a hardware budget beyond the checkpoint alone. The hosted API allows up to 128 questions with input under 256K tokens. The downloadable implementation defaults to a 262,144-token sequence limit, but the cards say training used sequences up to roughly 32K tokens and longer inputs have only been spot-checked. Teams planning to classify long documents should evaluate their actual document lengths and evidence locations before relying on the maximum context setting. Perplexity also discloses that training includes official training splits from some Decision Index benchmarks, with test-set decontamination. Its reported benchmark improvements are useful context, but a private evaluation set should decide whether the model fits a particular application. The release gives developers a cheaper Flash option and more efficient shared-context decisions. The strongest reason to try it is a measurable improvement in a specific workflow. Keep model selection, question grouping and operating thresholds together in that evaluation, because each can change which decisions the application makes. I'm Paras. I work at the intersection of AI, engineering and distribution. I help technical founders turn model launches and engineering tradeoffs into clear stories developers can understand and use. [Book a 30-minute call](https://cal.com/paras-madan/30min).