Perplexity Launches Fast Search in Search API
Perplexity announced the addition of fast search to its search API, with retrieval and sorting running on its self-developed Rust service, Photon. The company claims that a small team of engineers, along with hundreds of intelligent agents, built it, with 95% of search results returned within 230 milliseconds. This interface is aimed at developers, returning the search results themselves, not conversational answers.
The search API will be open to the public in September 2025, selling the web index behind the answering engine to external applications, covering hundreds of billions of pages, with a pay-per-use fee of about $5 per thousand queries, excluding large model tokens. Internal evaluations at that time showed a median latency of 358 milliseconds, with p95 below 800 milliseconds, compared to Exa's median of about 1375 milliseconds, Brave's about 513 milliseconds, and SERP scraping represented by Tavily at about 1342 milliseconds. Daily query volume is approximately 200 million. Fast search reduces p95 from below 800 milliseconds to 230 milliseconds, targeting every millisecond in multi-turn retrieval by intelligent agents that contributes to the first token time.
Two weeks ago, Perplexity announced Cobbledb: a key-value hot storage written in about 40,000 lines of Rust, replacing DynamoDB, with median batch read latency dropping from 31.4 milliseconds to 5.60 milliseconds, a reduction of about 82%, and p99 dropping from 123 milliseconds to 24.2 milliseconds, claiming storage layer costs could decrease by more than 20%, while still using RocksDB underneath. Photon and Cobbledb are both self-developed service layers, one managing retrieval and sorting, the other managing hot reads. In the documentation, answer interfaces like Sonar have already divided search_type into fast and pro, with simple facts taking the fast path and complex research going through pro, charging based on trigger types.
The product positioning is to directly hand over indexing, hybrid retrieval, and multi-segment sorting to RAG and intelligent agents, rather than adding another layer of generation. The interface supports domain and language filtering, time windows, and a maximum of five queries per single request. After Microsoft shut down the Bing Search API, alternatives for developers include Brave, Exa, and various SERP proxies. Perplexity differentiates itself by using its own index and fine-grained fragment scoring, rather than scraping others' result pages.
Mechanically, it sells to application developers who want to embed web retrieval into their own model loops, while selling higher-latency third-party retrieval and SERP scraping. The event-driven approach is to mark internal latency optimizations externally as tiers in the API product line. Funding flows from per-query retrieval fees to self-developed Rust services and index updates; benefiting are applications that can write 230 milliseconds into intelligent agent orchestration and Perplexity's developer revenue, while the median remains above 500 milliseconds for the search API. Hundreds of intelligent agents are building infrastructure, shifting human effort from stacking features to stacking latency percentiles.
Source: Public Information
ABAB AI Insight
Perplexity started with the answering engine, first selling the generative Sonar API in early 2025, and then separating pure retrieval without models in September of the same year to avoid binding developers to its own generation layer. A median of 358 milliseconds is already faster than Exa, and a p95 of 230 milliseconds pushes retrieval from "acceptable network round trip" to the range of "multiple hits in a single inference." A small team plus hundreds of intelligent agents built Photon, which aligns with the narrative of replacing DynamoDB with intelligent agents and Rust two weeks ago: using code generation to compress infrastructure iteration rather than expanding the retrieval team.
The capital path is "consumer-side answers support the index, developer APIs spread fixed costs." With 200 million daily queries, crawling and indexing become sunk costs, and the search API recoups through per-thousand charges; the fast tier exchanges lower latency for higher call density, as intelligent agents may perform dozens of retrievals in a single task. Cobbledb cuts hot reads, Photon cuts retrieval and sorting, both avoiding cloud vendors' byte-based billing. After Microsoft shut down the Bing API, there is no default platform retrieval at this layer, with Brave, Exa, Tavily, and Perplexity competing for the same RAG budget.
Analogous to how Google turned custom search into an API, OpenAI treats retrieval as a tool, and Cloudflare and Fastly replace general cloud databases with self-developed edge solutions: all are productizing internal latency advantages. The difference is that Perplexity sells both generation and retrieval, and fast search prevents developers from solely using its answers while outsourcing retrieval to others. The industry phase has shifted from "Is there real-time web access?" to "Can retrieval in a single intelligent agent loop be less than the model's first token?" Whoever first brings p95 down to just over 200 milliseconds will be more suitable as the default tool.
Structural judgment belongs to technological substitution. Pricing power has shifted from "Whose result page can be scraped?" to "Whose index can return citable fragments in the 200-millisecond range?" The mechanism is that large model first tokens can take seconds, so retrieval must be an order of magnitude shorter than generation; otherwise, multiple intelligent agent loops will be dragged down by retrieval. Rust hot paths and self-developed storage remove tail latency from cloud database billing. Intelligent agents build infrastructure, simply accelerating this replacement.
ABAB News · Cognitive Laws
In the intelligent agent loop, retrieval must be an order of magnitude shorter than generation.
Self-developed storage and self-developed retrieval are targeting the same cloud bill.
After shutting down the platform search API, latency numbers become the new shelf positions.