Perplexity Decider v1.3 Hits 91.6% Accuracy

Why this matters

Perplexity is betting that a lot of agent plumbing doesn't need a chatbot at all, just a model that scores probabilities and hands back exact text spans. Locate is the tell here: if a decision model can pull the right substring from a contract or a product page without hallucinating around it, that's a cheaper, more auditable replacement for a bunch of regex-and-LLM glue code teams are running today. For builders, the Apache 2.0 weights on Hugging Face matter more than the headline number. You can actually test the 91.6% claim against your own classification tasks instead of trusting a vendor benchmark.

We'd treat the self-reported figures as a starting point, not a verdict. "6.54 points above production classifiers" and "up to 12.9 points in finance" are Perplexity's own comparisons, against its own baselines, on its own task set. That's worth watching once independent evals show up, especially on domain classification where the gaps are largest.

If Locate holds up outside Perplexity's test suite, it's a genuinely useful primitive for retrieval pipelines. If it doesn't, it's another benchmark that reads better than it runs.