Perplexity's Fast Search runs on a Rust engine written by a small team plus hundreds of agents for about $300K in tokens

Perplexity launched Fast Search in its Search API on September 24. It runs on Photon, a new Rust retrieval and ranking service that replaces a modified open-source engine. The numbers: 95% of results return in 230ms or under, 160ms median, internal p99 down from about 800ms to 65ms on 20% fewer machines. Pricing is $1 per 1,000 requests against $5 for standard search, with a reported 68% lower cost per task across six agentic benchmarks. (Perplexity community forum)

Two separate things here and both are useful.

First, the search call itself. An agent doing research fires dozens of searches per task, and at $5 per 1,000 the retrieval line item stops being rounding error. A 5x price cut with a 65ms p99 changes what search-heavy agent loops you can afford to run. I'd read the 68% cost-per-task figure as Perplexity's own benchmark, not an independent one, but the latency numbers are the kind that either hold up in your own timing or don't within an hour of trying.

Second, the provenance. A core retrieval engine, in Rust, written by a small team plus hundreds of agents, for about $300K in tokens. That's a dollar figure on agent-written infrastructure at the layer where correctness actually matters. Not a demo app. A ranking service with a p99 SLO.

Put that against the smallest version of the same claim. Dan Greenheck built TideWater, a browser island with birds, crabs, whales, wind-swayed cloth and a sailable boat, in about 8 hours using Opus 5.5 and subagents, and posted the receipt: $1,874.40 of tokens, 59% of his weekly usage. (dgreenheck/tidewater, created September 23, 528 stars by the 25th.) One person, one day, under two grand.

Klaviyo sits in the middle. 512 employees across legal, marketing, HR and engineering shipped 356 live internal apps in two weeks through K:Forge, which turns a Slack request into a deployed app in under three minutes, and 196 of those apps have their own databases. What makes it work isn't the generation, it's the fence: every app behind Okta SSO, private by default, Vercel Secure Compute for private network paths to internal databases, Wiz scanning everything. (Vercel)

$1,874 for a day of 3D work. $300K for a production search engine. Three minutes and an SSO gate for an internal tool. The cost of building software now has a price list, and the interesting column is what you spent on the guardrails.