Perplexity has launched Photon, a retrieval engine built on Rust designed to optimize search performance. The company reports that the new engine reduces p99 latency from 800 ms to 65 ms. Photon is available as a hosted API, and users can access it by setting the search_type parameter to fast in their POST requests.
The service is priced at $1 per 1,000 requests. Perplexity notes that Photon is not open source and cannot be self-hosted. The engine architecture utilizes a load balancer that routes requests to a broker, which then distributes tasks to shard groups for retrieval and ranking processes.
Internal testing across 3,554 tasks showed that the Fast search mode achieved a score of 64.3% with an estimated cost of $59.73. In comparison, the default preset scored 64.0% at a cost of $187.60, making the new option approximately 68% cheaper for users.
The company highlights a trade-off in search quality, noting that relevance scores on long-tail benchmarks decreased from 2.45 to 2.21. Answer availability also saw a decline of 2.9 percentage points. Perplexity suggests using the Fast mode for agentic workflows and the default for complex queries.
