Perplexity Decider v1.1: Open 27B Decision Model at Half Price

A Point Release, Days After Launch

Perplexity has updated its decision model less than a week after the first version went live. The original pplx-decider-v1-27b powered the company's Decisions API, which opened on October 1, 2026.4 On October 7, Perplexity released pplx-decider-v1.1-27b as an open-weight multimodal model, with the API available the same day and a listing on OpenRouter.2

The main change for buyers is price. Input now costs $0.02 per million tokens, half of v1's $0.04, and output tokens remain free.124 Calling v1 "obsolete" overstates it, but anyone who wired up v1 last week now has a cheaper default to switch to.

What a "Decider" Actually Does

The decider is not a chatbot. It chooses among options the developer defines and returns a probability for each one, rather than writing free text.1 At launch, the API answered in three typed formats: yes/no, multiple choice, and ordered score.4

That makes it a classification and judgment tool. Likely uses include grading retrieval results, routing requests, flagging hallucinations, or acting as an automated judge inside agent pipelines. The value depends on whether those probabilities are calibrated, meaning a 0.8 should be right about 80% of the time.

Both versions are 27B-parameter fine-tunes of Qwen3.8-27B, released on Hugging Face under the Apache 2.0 license.14 Running v1.1 yourself takes roughly 49GB of GPU memory, so it fits on a single high-end accelerator instead of a cluster.1 V1.1 accepts both text and image input.12

The context window is described two ways. One account rounds it to 250K tokens.1 The v1 documentation and an integration config both list 262,144 tokens.34 This is almost certainly the same limit stated with different precision. Latency rises with input length: v1's docs show responses under two seconds for a few hundred tokens, stretching to about 23 seconds near the context ceiling.4

The Benchmark Claims

Perplexity is presenting v1.1 as the leader in a young category. The company says it ranks first on Hugging Face's newly launched Decision Index (version 0.3).12 On October 9, CEO Srinivas added that it also tops the open DecisionBench benchmark at the lowest cost.2

The v1 numbers came from Perplexity's own evaluation: an 11-benchmark panel of 7,210 samples. There, pplx-decider scored 85.71% against 84.51% for a competitor called Jev.4 The largest gap was on RAGTruth, a hallucination-detection test, at 88.80% versus 77.27%.4 Outside that one result, the overall margin is narrow.

These figures need context. The 11-test panel was built by the vendor. The Decision Index is new and still at a sub-1.0 version number, so its methodology is unsettled. Leading a leaderboard that was just created says less than leading an established one. Because the weights are open, independent testers can check these claims without depending on Perplexity's API, and they should.

Where the Ecosystem Lags

A GitHub issue in the oh-my-pi project shows a practical problem with the new model type.3 A user registered Perplexity's decider through the Requesty router using a standard OpenAI-style completions setup and assigned it as a "judge." The tool then sent ordinary streaming chat requests instead of using Requesty's Decisions protocol.3

The decider rejected those requests. The tool then fell back to a prompted answer with synthetic probabilities of exactly 0 or 1.3 The issue's author warns that this output could be mistaken for real calibrated decision scores.3

This matters because the product's main selling point is calibrated probabilities. If a framework quietly replaces them with fake binary values, developers may build pipelines on confidence scores that mean nothing, and nothing will alert them. The issue also lists v1's $0.04 input price, which shows how fast third-party configs can go stale when a vendor reprices within a week.3

Reading the Move

The quick v1.1 release looks less like fixing a broken product and more like a land grab in a category Perplexity wants to define. It combines open weights, a permissive license, near-zero input pricing, and free output. That makes the decider an easy default for anyone who needs cheap, structured judgments at scale.

The open questions are not mainly about model quality. The first is whether the benchmarks hold up under independent testing. The second is whether routers, agent frameworks, and SDKs add native support for decision-style endpoints. Until that tooling exists, the calibrated probabilities that set this model apart can be lost in translation, and developers should confirm they are receiving native decision output.