Perplexity Decider v1.3 is the better choice than TypeSafe AI's Jev 1.14 for most classification workloads. On the vendor's own benchmarks it leads on accuracy, robustness and price, and it ships as Apache-licensed weights you can self-host. Perplexity reports a cost of $26.47 per million requests for Decider v1.3 against $68.10 for Jev 1.14, and $13.24 for the distilled Decider Flash v1.3 (Perplexity). The "one fifth" price gap applies to Flash, not the full model, and the real-world accuracy margin over Jev is small.

TL;DR

  • Perplexity released Decider v1.3 and Decider Flash v1.3 on 12 November 2026 under the Apache license, on Hugging Face, OpenRouter and the Perplexity Decisions API.
  • Vendor-reported accuracy: 91.6% mean across 63 query-classification tasks, 6.54 percentage points above Perplexity's own production classifiers, and ahead of Jev 1.14 (Perplexity).
  • Vendor-reported cost for the same 1M-request workload: $26.47 (v1.3), $13.24 (Flash), $68.10 (Jev 1.14).
  • Robustness is the clearest gap: v1.3 changed its answer on 0.27% of reshuffled multiple-choice questions, against 2.69% to 4.84% for Jev 1.14 (Perplexity).
  • On real production traffic the judge-agreement gap over Jev is only 0.63 points, so the strongest reasons to switch are the open weights and the new "locate" feature.
  • Last verified: 14 November 2026. All headline figures come from Perplexity and have not been independently reproduced.

What is Perplexity Decider v1.3?

Decider v1.3 is a decision model: a classifier that takes one shared "state" plus several questions and returns calibrated probabilities in a single forward pass, without generating tokens. It handles yes/no, multiple-choice and ordered score questions, and suits judging outputs, routing, moderation, tagging and reranking. It is a successor to Decider v1, v1.1 and v1.2, and Perplexity says it was inspired by the "System One Model" idea behind Jev from TypeSafe AI (Perplexity).

The large model starts from Qwen3.8-27B. Flash is distilled from a Qwen3.5-4B base. If you are weighing Jev against general-purpose LLM APIs first, our Jev vs Claude cost verdict covers that side.

Is Perplexity Decider v1.3 better than Jev 1.14?

Yes on every benchmark Perplexity published, though the size of the lead varies a lot. The table summarises the vendor-reported figures (Perplexity); independent trade coverage of the release recorded the same headline claims.

Measure Decider v1.3 Decider Flash v1.3 Jev 1.14
Cost per 1M requests $26.47 $13.24 $68.10
Multilingual mean (19 tasks, 8 languages) 88.16% 86.50% 85.41%
Vision suite mean (5 tasks) 76.99% 72.77% about 75.5% (v1.3 is +1.47 pp)
Production judge agreement 93.98% not reported 93.35%
Answer changes under option reshuffling 0.27% not reported 2.69% to 4.84%
Answer changes on repeated identical input none not reported 1.88%

Two details matter. On the seven public-dataset domain sweeps (claim verification, healthcare, cybersecurity, programming, government, law, finance), v1.3 leads in all seven, by up to 12.9 points in finance and 9.3 in law (Perplexity). Flash, however, trails Jev 1.14 in finance, programming and law. If those domains are your workload, Flash is not a free swap.

How much does Decider v1.3 cost per request?

Dividing Perplexity's per-million figures (reported for the same one-million-request workload (Perplexity)) gives about $0.0265 per 1,000 requests for v1.3, $0.0132 for Flash and $0.0681 for Jev 1.14, so Jev costs roughly 2.6x as much as v1.3 and 5.1x as much as Flash on the traffic Perplexity measured (Perplexity). Larger general-purpose models in the same comparison cost far more: $312.21 for GLM-5.3-Flash and $349.48 for DeepSeek V4.1 Flash on the same workload (Perplexity).

Perplexity itself flags the caveats. The workload is specific to its traffic, hardware-cost accounting for locally run Decider is not specified, and there is no matched latency comparison. Treat the dollar figures as a ratio to test, not a quote.

What is new in Decider v1.3?

Three changes stand out, all described in the Perplexity announcement:

  • Isolated questions. Each question now attends only to the shared state and its own tokens, which Perplexity says gives more stable results than the everything-sees-everything pattern in v1.2.
  • Locate. The model can return the exact matching text span with start and end character positions, covering named entity recognition, PII detection and extractive fields such as invoice totals. On 17 out-of-domain tasks, v1.3 averaged 69.0 and Flash 63.6, against 43.3 for Liquid d1, 41.5 for Cloudflare Clairic and 33.6 for GLiNER large v2.5 (Perplexity).
  • Refined training. A three-stage recipe (generalist fine-tuning, LoRA refinement, a LoRA vision step) with weights averaged over three seeds.

Locate adapters stay unmerged and are activated only for locate requests, because merging them reduced decision accuracy.

Should you switch from Jev 1.14 to Decider v1.3?

Switch if you want to self-host, need span extraction, or care about stable outputs. Stay if Jev is already integrated and your volume is low enough that the price difference is immaterial.

The case for caution is the production sample. Across 2,318,887 questions drawn from 1 million deduplicated, PII-masked Decisions API requests, judge agreement was 93.98% for v1.3, 93.68% for v1.2 and 93.35% for Jev 1.14 (Perplexity). A 0.63-point gap will not transform a pipeline. The open Apache weights, absence of API lock-in and the robustness numbers are the stronger arguments. This pattern of a small score gap with a large licensing gap also shows up in other open releases, such as LongCat 2.5 under MIT.

How do you upgrade or get started?

Perplexity says the pplx_decider interface is unchanged in shape from v1.2, with the same state and questions dictionary, and that the Decisions API needs no payload changes. OpenRouter serves both models. For self-hosting, download the checkpoints from the Hugging Face collection and follow the model card.

A minimal sketch of the flow, to be checked against the model card because exact argument names may differ:

from pplx_decider import Decider # name per model card; verify model = Decider.from_pretrained("pplx-decider-v1.3", device="cuda") result = model.decide( state="Customer message: I was charged twice for order 4471.", questions={ "is_billing_issue": "yes_no", "urgency": "score", }, ) # result holds calibrated probabilities per question 

Then tune your decision thresholds on a small labelled dev set rather than using fixed cutoffs, which is Perplexity's own advice. If you use locate, test both the decision path and the locate path after upgrading, since the adapters are activated separately.

What are the alternatives to Jev and Decider?

Other open options exist. Nace AI's Drex models (Apache, with weights and inference code) report 67.52 on Decision Index 0.4.1 against 60.70 for Jev 1.14.0 (arXiv 2611.07854). That is another vendor-reported number, so run your own evaluation. For coding-agent workflows, a decision model can gate steps such as whether to spawn work in the background, a choice we compare in OpenCode background subagents vs inline, and can score plans before execution as in plan-first coding with shadcn improve. For the wider open-source agent field, see the DeepSeek harness leaderboard.

FAQ

Q: What is Perplexity Decider v1.3?
A: It is an open-weights decision model from Perplexity that answers yes/no, multiple-choice and score questions about a shared input in one forward pass, returning calibrated probabilities. It was released on 12 November 2026 under the Apache license.

Q: Is Decider v1.3 really one fifth the price of Jev 1.14?
A: Only Decider Flash v1.3 is close to that. Perplexity reports $13.24 per million requests for Flash and $26.47 for v1.3, against $68.10 for Jev 1.14, so the full model costs about 39% of Jev (Perplexity).

Q: Can I run Decider v1.3 locally?
A: Yes. The weights are on Hugging Face under the Apache license, and the pplx_decider interface offers from_pretrained and decide. Perplexity does not specify hardware costs for local runs, so budget for your own GPU.

Q: What does the locate feature do?
A: It returns the exact text span that answers a question, with start and end character positions. Uses include entity recognition, PII detection and pulling fields such as NDA parties or invoice totals.

Q: Are the benchmark results independent?
A: No. The figures come from Perplexity's own evaluations, and the cost comparison carries stated caveats on workload, hardware and latency. Run a small test on your own labelled data before migrating.

Last verified: 14 November 2026. Corrections: none to date. This article was prepared with AI assistance and checked against the primary source linked above; our approach is described on the How We Work page.