Perplexity, the AI search company, has released a Context API that lets an agent submit a list of up to 10 URLs and receive the readable content of each page. The pitch, laid out on the Perplexity blog, is that search snippets are too thin for real work and raw HTML is too noisy. The new endpoint sits between the two.

Developers choose the output per request. Markdown is the cleaned page text, HTML is there for teams that parse their own, and links come back as absolute URLs, deduplicated and in page order. The fourth format is snippets: the passages most relevant to a query, trimmed to a token limit the developer sets for each page. Perplexity says these are extracted, not summarized, so the model reads the page’s original wording.

Every successful result also carries metadata: title, description, language, author, published date and Open Graph tags. To produce the Markdown, Perplexity says a model scores each block of a page as main content or boilerplate and keeps only the former. Results arrive in the same order as the URLs were sent.

Freshness is the control developers will touch most. Out of the box, the API looks in Perplexity’s index for a stored copy no more than two days old, and reads the live page only when none exists. Setting max_age_seconds to 0 forces a live read, which suits volatile pages such as pricing, scoreboards and news as it breaks. A higher value suits slow-moving reference material, and cache_only set to true guarantees no live request is ever made. Each result reports whether it came from storage or a live fetch, and when.

Batches fail gracefully. The URLs run concurrently, and every one gets a separate clock and a separate status code, so a single slow page does not sink the other nine. Only the timeouts need a second attempt.

Perplexity names four jobs where a snippet falls short: writing an API integration from documentation, comparing current prices, answering return-policy questions, and researching across several sources. The pattern repeats. The excerpt gives the headline fact and drops the exception, the billing condition or the methodology that changes the answer.

The benchmark numbers deserve a careful read. Perplexity tested its API against Firecrawl, Exa, Parallel, Keenable and Context.dev using thousands of addresses drawn from genuine agent traffic, spanning 715 sites that include government, academic and non-English pages. A judge model scored outputs without knowing the provider. Perplexity reports that its API produced real page content for 92.3 percent of URLs, recovered 87.9 percent of key facts and reached 80.3 percent content precision. Perplexity designed the test, chose the judge and published the chart. No independent replication accompanies it, and a failed fetch scores zero on recall and precision, so reliability drives all three figures at once.

Pricing is simple. The API costs $1 per 1,000 URLs fetched, and only URLs that return an ok status are billed. At list price, a full batch of 10 pages costs one cent, and a page that errors out costs nothing.

At a tenth of a cent per page, the build-versus-buy question for a page-reading step starts with one number you probably have not measured: what share of your own agent’s fetches fail or return cookie banners today. Run 100 of your real URLs through the endpoint before taking Perplexity’s 80.3 percent as your own.

Perplexity, launch announcement for the Context API on the Perplexity blog (perplexity.ai).