Perplexity Decisions API: Open 27B Decider Model Takes On Jev

What Perplexity shipped

On October 1, 2026, Perplexity launched a Decisions API and released the weights of the model behind it, pplx-decider-v1-27b, under the permissive Apache 2.0 license.12 The release is unusual in what the model is built to produce. It does not generate free-form text. Instead, it returns a probability distribution over a fixed set of possible answers.1 The model is multimodal.1 It is a fine-tune of Qwen3.8-27B and is hosted on Hugging Face, with a context window of 262,144 tokens.2

The API supports three typed response formats: yes/no, multiple choice, and an ordered score.2 Perplexity charges $0.04 per million input tokens, and output is free.2 According to the documentation, a request with a few hundred input tokens comes back in under two seconds. Latency rises to about 23 seconds when a request approaches the 262,144-token input limit.2

The timing also stands out. Perplexity made the announcement only hours after Cloudflare released Clef.1 That suggests this category of product, models built for structured judgments rather than chat, is drawing attention from several companies at the same moment.

The benchmark claim, and its limits

Perplexity's main competitive claim is that pplx-decider beats a rival decision model called Jev. The reported scores are 85.71% for pplx-decider and 84.51% for Jev.12 Those figures come from a panel Perplexity assembled itself, made up of 11 benchmarks and 7,210 samples.2 The largest gap is on RAGTruth, a test associated with spotting hallucinations in retrieval-augmented generation. There, pplx-decider scored 88.80% against Jev's 77.27%.2

Two points deserve attention.

• The overall lead is narrow. A difference of roughly 1.2 percentage points across a panel of this size is a modest advantage, not a decisive one.
• The averages are uneven underneath. An 11.5-point win on RAGTruth combined with a 1.2-point overall margin implies that Jev is close to pplx-decider, or ahead of it, on at least some of the other ten tests. The available reporting does not break those results out.

The panel is also Perplexity's own selection, and vendors who design their own evaluations tend to choose tests that favor their models.2 Because the weights are open, though, independent researchers can rerun the comparison, and those results will matter more than the launch figures.

Why a "decision model" matters

A model that outputs a probability distribution over set answers addresses a common practical problem. Many production AI systems use large language models as classifiers or judges. They ask questions such as whether a response is grounded in its sources, whether a document is relevant, or which of several options is best. Getting those answers by parsing generated text is slow, costly, and fragile.

Typed outputs paired with calibrated probabilities make this kind of judgment something a developer can call reliably and set thresholds on.12 A developer could, for example, route only low-confidence cases to a human or to a more expensive model.

The pricing supports this reading. Charging only for input and making output free fits a workload where the response is a short label or score and nearly all the cost is in reading the material.2 The long context window and the published latency curve indicate that Perplexity expects customers to submit large inputs, such as whole documents or long retrieval results, and receive a single verdict.2 The RAGTruth result points the same way. Checking whether a generated answer is supported by retrieved sources is central to Perplexity's own answer-engine business, so this is plausibly where the company's internal needs and its product strategy meet.

The open-weights angle

Releasing the model under Apache 2.0 while selling a hosted API is a familiar two-track approach.12 Teams that want to self-host, fine-tune further, or keep data on their own infrastructure can do that. Others can pay a low per-token rate and avoid running a 27B model themselves.

For Perplexity, the open release also builds credibility and adoption in a category that rival launches are now crowding.1 Basing the model on Qwen3.8-27B shows how much of the current open-model ecosystem is built on top of strong open base models.2

The takeaway

The launch is more significant for its format than for its scores. The benchmark lead over Jev is real but small and self-reported, and it should be treated as provisional until independent evaluations appear. The more lasting contribution is a cheap, open, long-context model designed specifically to make structured judgments with probabilities attached.

If independent testing confirms the RAGTruth advantage, pplx-decider could become a default option for grounding checks in retrieval pipelines. Even if it does not, the near-simultaneous arrival of Perplexity's and Cloudflare's offerings suggests that "decision APIs" are becoming a distinct product category.1