Perplexity Decider v1.1 Halves Price to $0.02 Days After Launch

A price cut before the ink dried

On October 6, 2026, Perplexity released pplx-decider-v1.1-27b, an updated version of its open-weight decision model. It is priced at $0.02 per million input tokens, half the rate of the original 2. The company says the new checkpoint scores highest on Hugging Face's new Decision Index 0.3 benchmark. It also keeps text and image input and offers roughly a 250k-token context window 2.

The update came just five days after the first version shipped. On October 1, Perplexity launched its Decisions API alongside Apache-2.0 weights for pplx-decider-v1-27b. That model is a fine-tune of Qwen3.8-27B, and it returns a probability distribution over a fixed set of answers rather than generated text 15. The original pricing was $0.04 per million input tokens, with output tokens free and no per-request fee 13. Halving that so quickly says a lot about how crowded this niche has become.

What a "decision model" actually is

The pitch is narrow on purpose. You send a state, which can be text, JSON or images, along with one or more questions. The model answers each question in one of three typed shapes: a yes/no-style answer, a multiple-choice pick, or an ordered score. Each answer comes with probabilities and a confidence value instead of prose 35.

To build this, Perplexity removed the model's language-generation head and replaced it with a 255-option readout. That means the released weights cannot write text at all 4. A single forward pass scores every allowed answer 1. The API accepts up to 262,144 input tokens and between 1 and 128 questions per request 13.

Perplexity was not first here. TypeSafe's Jev arrived September 15. OpenAI has announced its own Decisions API, and Cloudflare's Clef landed only hours before Perplexity's launch. That made pplx-decider the third Jev-style model in about two days 1.

The benchmark claim, and why to read it carefully

Perplexity's headline number is 85.71% overall versus 84.51% for Jev. That result comes from an 11-benchmark, 7,210-sample panel the company assembled and ran itself 15. Some individual gaps are large:

• WinoGrande: 90.70% vs. 83.30% 1
• BBH: 94.27% vs. 82.80% 1
• RAGTruth, the widest gap: 88.80% vs. 77.27% 5

The aggregate lead, however, is only about 1.2 points. Several per-benchmark wins are much wider than that. Read together, this suggests Jev likely does better on some of the remaining tests, which narrows the overall margin. A self-selected panel, a small aggregate edge, and mixed underlying results are exactly the conditions under which a vendor's "we beat the competitor" claim deserves skepticism.

The v1.1 announcement shifts the argument to a third-party venue, Hugging Face's Decision Index 0.3 2. That is a step toward independent comparison. Still, the claim comes from Perplexity's own post, and the index is newly versioned.

Early real-world friction

The community responded quickly. Within days, AveLabs published "Perplexing-Pegasus," which grafts Qwen3.8-27B's original language-model and MTP heads back onto the decider's fine-tuned backbone. The result is one set of weights that can both answer typed decisions and chat 4. In AveLabs' tests across 2,196 items, its NVFP4 build matched the original BF16 decider at 76.0% overall and gave the same answer 96.7% of the time 4. Its numbers on JevBench public (87.4%) and WANLI (68.1%) also give outsiders an independent reference point for the original model.

Latency is a rougher spot. Perplexity's docs show responses under two seconds for a few hundred input tokens, rising to about 23 seconds near the context ceiling 5. One developer replying to the v1.1 announcement said they could not run 10 questions simultaneously in practice, and that batching questions increased per-question latency 2. That complaint sits uneasily next to the documented limit of up to 128 questions per request 1. The gap between spec and experience is something Perplexity has not addressed publicly.

The takeaway

The benchmark lead is the least convincing part of this story. The price cut is the most telling part. Within a single week, four players either shipped or announced near-identical products. Output is free, and inputs now cost two cents per million tokens. In that setting, decision models are commoditizing almost as soon as they appear.

Perplexity's Apache-2.0 release is notable, as the fast community remix shows 4. The 50% cut also signals that the company expects to compete on cost and openness rather than on a narrow, self-reported accuracy margin. Developers should benchmark these models on their own tasks, and they should test latency at realistic batch sizes before committing.