Perplexity’s pplx-decider-flash is a smaller model for returning decisions and probabilities, not writing chat replies. Its headline API price is half that of the company’s larger Decider model, but the saving disappears once you ask two questions about the same input, according to OrcaRouter’s analysis.
Perplexity Watch analysis
What happened
Perplexity’s Flash model is described as a 4B-class decision model that takes text, JSON or images and returns typed results such as yes/no probabilities, choices or scores. The analysis says the API charges $0.01 per million input tokens, with no separate output-token charge.
The catch is how it handles multiple questions. Flash evaluates each question independently, billing the input state again each time. On a 10,000-token state, the article’s calculation puts one question at $0.0001, two at $0.0002 and three at $0.0003. The larger model, at twice Flash’s input rate, costs $0.0002 for the same state whether asked one, two or three questions.
OrcaRouter also reports Perplexity’s benchmark results: Flash scored 60.05 on the company’s Decision Index when questions were asked one at a time, and 58.88 when grouped. The larger model scored 68.09 on grouped questions. These are Perplexity’s own figures, not an independent comparison.
Why it matters
Flash is aimed at a useful but specific job: a low-cost gate that classifies a request or chooses a tool before a larger model does the generating. For a single question, it may offer a cheaper way to make that decision. For several questions about the same material, the repeated input billing can make the bigger model less expensive. “Smaller and cheaper” needs an asterisk shaped like your workload.
Our read
The practical lesson is to count questions per input state before reaching for the bargain model. If your system asks just one, Flash’s pricing may fit. If it routinely asks two or more, run the arithmetic against the larger model before calling it a saving. And treat the benchmark gap as a vendor-reported starting point, not a settled leaderboard.
What to watch
- Whether Perplexity publishes independent evaluations or a fuller breakdown of model errors.
- Whether the API price or the versioned model changes.
- How Flash performs on real workloads with multiple questions per input state.
Discussion spark: Would you use a smaller decision model for a one-question gate, or is the extra benchmark performance of a larger model worth the cost?
Sources and evidence
- pplx-decider-flash: What Perplexity's Quiet 4B Judge Ships – OrcaRouter (3 November 2026, 22:30 UTC)
not affiliated with or endorsed by Perplexity