Perplexity has released two Apache 2.0-licensed models designed to classify, score or choose, rather than write long answers. The smaller model starts at $0.01 per million input tokens, giving developers a low-cost option for structured decisions in AI applications.
Perplexity Watch analysis
What happened
The models, pplx-decider-v1.2 and pplx-decider-flash-v1.2, have 27 billion and 4 billion parameters respectively, according to TokenPost. The larger model costs $0.02 per million input tokens and the smaller $0.01; output tokens are free. Both are available through the Perplexity API and Vercel AI Gateway. Their repositories are private and require authorised access, despite the Apache 2.0 licence.
They process text and images to return classifications, choices or scores. Perplexity’s documented uses include routing agent tasks, scoring, labelling and filtering. Developers can also send several questions about the same text or image in one request, potentially reducing the number of inference passes. The report cautions that combining questions can reduce accuracy with the 4B model, whose default format therefore uses one question per request.
TokenPost reports that the 27B model scored 67.75 on Decision Index 0.3 using one question per request, up from 62.30 for its predecessor in the same format. With joint requests, it scored 68.09. The article says some official training splits for evaluated tasks were included in v1.2’s training data and decontaminated against test sets. These are reported benchmark results, not a guarantee of performance on a developer’s own workload.
Why it matters
A model that returns a defined choice can fit neatly into software workflows where a free-form paragraph is more decoration than help. Task routing, labelling and filtering are familiar jobs, but developers still need to know whether a model is accurate enough for the particular decision and what it costs at scale. The low input prices make experimentation easier; the model’s limits and evaluation details still matter more than a bargain headline.
The licence and access arrangements also deserve a closer look. Apache 2.0 is permissive, but the repositories are private and require authorisation, so “open-licensed” does not mean anyone can simply download the weights today.
Our read
This is a practical addition to the growing menu of specialised AI tools: smaller models for constrained outputs, rather than asking a general chatbot to narrate every decision. The useful test is not whether a benchmark score looks tidy, but whether the model handles your real inputs reliably and cheaply. And if you were hoping “Apache 2.0” meant immediate access to the weights, there is a gate to clear first.
What to watch
- Whether the model weights become broadly accessible and what authorisation requires.
- How accuracy holds up on real-world classification, routing and scoring tasks.
- Whether the shared-input format saves enough inference work to justify any accuracy trade-off.
Discussion spark: For routine software decisions such as routing and labelling, would you trust a small specialist model, or keep those choices in conventional code?
Sources and evidence
- Watch Perplexity official source (17 October 2026, 11:00 UTC)
not affiliated with or endorsed by Perplexity