Perplexity's Fast Search gets AI agents web results before you can say 'let me Google that'

Perplexity's new Fast Search mode, built on its in-house Photon engine, returns 95% of web results in 230 milliseconds or less for US$1 per 1,000 requests. It's aimed at AI agents that search the web many times per task.

Not every piece of AI news is about a chatbot. Some of it is plumbing, and this week's plumbing is pretty good. Perplexity has launched Fast Search, a mode in its Search API built for AI agents, the programs that might search the web dozens of times to finish one task. It runs on Photon, a retrieval and ranking engine Perplexity built in-house using the Rust programming language. Perplexity says Fast Search returns 95% of results in 230 milliseconds or less, with a median of 160 milliseconds.

Photon now handles all of Perplexity's production search traffic. According to MarkTechPost, it cut the slowest-end retrieval time from around 800 milliseconds to around 65, on roughly 20% fewer servers. Fast Search costs US$1 per 1,000 requests, and developers turn it on by setting the search type to 'fast' in a search request. There is a trade-off: relevance dips slightly on Perplexity's own benchmark, though it reports agent tasks cost about 68% less than on the default setting, at similar quality.

Most businesses won't touch this directly. You'll feel it through the tools you use, or the ones a developer builds for you. If you've commissioned an agent that researches suppliers, checks competitor prices or pulls news for a weekly brief, search speed and cost add up quickly across dozens of lookups per job. It's the difference between a courier who stops at every set of lights and one who knows the back streets.

My advice: if you run an agent that searches the web, ask whoever built it which search service it uses and what each lookup costs. For high-volume, low-stakes jobs like news monitoring, a faster, cheaper mode is worth testing. For work where accuracy matters more than speed, such as compliance research or pricing decisions, keep the standard mode and compare answers side by side before you switch anything.

Faster, cheaper search won't make the evening news, and that's fine. The best infrastructure is the kind you only notice when the invoice gets smaller.