Perplexity Cuts Search API Prices for AI Agents With "Fast Search"

Perplexity has added a cheaper, faster tier to its Search API, aimed at developers whose AI agents query the web over and over in a single task. The company announced Fast Search and the engine underneath it, Photon, on September 24, 2026.

What launched

Fast Search is a new option inside the Perplexity Search API, which the company first opened to developers in September 2025. Developers choose it by setting search_type to "fast" on a request to the search endpoint. Leaving the setting out uses standard web search. Both return results in the same format, and Fast Search accepts between one and 20 results per request.

Perplexity's measurements put a Fast Search call at 160 milliseconds at the median and 230 milliseconds at the 95th percentile.

The price

Perplexity lists Fast Search at $1 per 1,000 successful requests, against $5 per 1,000 for standard search. At those rates, 100,000 requests cost $100 instead of $500.

Its pricing documentation says a successful request counts as one billing unit even when it bundles up to five queries, and that the Search API carries no token charges. Rate-limited, invalid, and failed requests are not billed. Inside Perplexity's separate Agent API, a Fast Search tool call is listed at $0.001, with model tokens billed separately.

Photon, the engine underneath

Perplexity says it previously adapted an open-source search engine and ran into limits on cost, slow responses at the high end, and the time it took to recover or add serving machines. Photon is its replacement for retrieval and ranking.

The company's reported results:

• 99th-percentile retrieval-and-ranking latency fell from about 800 milliseconds to 65 milliseconds. That is one internal stage, not an end-to-end API call.
• Index building now happens on separate machines, so serving nodes attach finished index versions instead of reindexing in place.
• Photon stores about 2.5 times as much data per document as the old system, giving ranking more to work with.
• It uses about 20 percent fewer equivalent serving machines than the old system's content nodes.

The trade-offs

Perplexity published the costs along with the gains.

On a set of 3,554 tasks across six agent benchmarks (WideSearch, BrowseComp, DSQA, FRAMES, SEAL-0, and SEAL-Hard), Fast Search scored 64.3 percent at an estimated total cost of $59.73, against 64.0 percent for default search at $187.60. That is the source of the company's "68 percent lower cost" figure, and it covers estimated model plus search cost for those tasks only.

On narrower retrieval tests of hard queries, coverage, and diversity, Fast Search did worse. Relevance fell from 2.45 to 2.21 on Perplexity's scale, and answer availability fell from 0.596 to 0.567. Perplexity's documentation recommends standard search for rare, difficult, or ambiguous questions.

Why this matters for AEO

Every AI answer engine and agent framework needs a search layer. When that layer gets five times cheaper, developers will call it more often, which means more automated systems reading the web on behalf of users.

For publishers and brands, two implications stand out:

• Speed favors clear pages. A fast retrieval tier that sometimes surfaces a weaker set of results rewards pages whose relevance is obvious from the title, headings, and opening paragraph.
• Agents read more sources per task. An agent that can afford 30 searches instead of five will consult more pages. That widens the set of sites that can be cited, if they are crawlable and easy to extract from.