The company's new 'Fast Search' API delivers a p95 latency of 230 milliseconds and a median p50 of 160 milliseconds for retrieval and ranking tasks. This service is powered by Photon, a new Rust-based engine that reduces the cost per task by 68% across six agentic benchmarks while maintaining comparable quality. Photon now handles all retrieval and ranking across the entire platform, replacing a previous open-source engine.

Perplexity built Photon using hundreds of AI agents and $300k in tokens, creating a system that uses about 20% fewer serving machines and stores 2.5x more data per document. The service is currently the default search for Nous Portal subscribers using the Hermes Agent. Internal testing indicates a slight trade-off for long-tail queries, which showed a 0.24 point drop in relevance and a 3% decrease in answer availability.