
Image: Perplexity
Why it mattersA team pointing an agent at the open web has to do this work themselves. Perplexity now sells it as one call at a priced rate.
An agent that reads the open web has to clean the page before the model sees it. Raw HTML carries menus, cookie banners, ads and tracking code. All of that confuses the model and runs up the token bill. Perplexity opened an API this week that does the cleaning in one call.
Perplexity launched the Context API on 17 November. The company says it reads up to ten URLs in one call. For each page it returns clean markdown, HTML, a list of links, and snippets that match a query. It replaces the step most agents do with a fetcher and an HTML parser.
What one call gives you
Each page comes back in four shapes. Markdown holds the main text without navigation or footer code. HTML returns the raw page for custom parsing. Links lists every absolute URL in reading order. Snippets picks passages that match a query the developer sends.
Each result also carries metadata. Perplexity names title, description, language, author, publish date and Open Graph tags as fields it extracts.
Freshness is a flag. By default the service returns a cached copy if the page was fetched in the last two days. A developer can set max_age_seconds to zero to force a live fetch. Setting cache_only accepts only a cached page. Each result says which path it came back on.
Price and vendor numbers
Perplexity says the Context API costs one dollar per one thousand URLs fetched. A failed URL is not billed.
The company also publishes its own benchmark against unnamed competitors on 1,000 URLs. Perplexity reports a 92.3 percent usable-content rate, 87.9 percent fact recall and 80.3 percent content precision. These are Perplexity's own numbers against its own test set. No third party has repeated them yet. Treat them as the vendor's claim.
A team that wants to compare a different extractor can run both on the same ten pages. The Context API returns the extraction and the metadata in the same JSON reply.
The Context API bills on top of the model that reads its output. Markdown is shorter than raw HTML. So the model-side token bill is smaller than scraping the whole page.
For a team building a GEO tool or an answer engine, the gain is simpler. Perplexity is both the search engine and the extractor. The same vendor that answers a query names the pages the answer came from. This API reads those pages. One pipeline covers both halves.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


