Perplexity’s Context API reads web pages for agents – $1 per 1,000 successful fetches
Perplexity’s new API takes a list of URLs and returns each page’s main content as Markdown, with its links and the passages relevant to a question. According to the company’s own comparison, the API returned usable content for 92.3 per cent of URLs.
Perplexity released the Context API on 17 November. The company’s Search API finds pages by keyword. The Context API reads pages whose URLs an AI agent already has. Search result snippets often miss essential details such as pricing terms, the company wrote in its blog. HTML fetched directly, by contrast, is full of menus, cookie notices and adverts.
To build the Markdown version, the API splits a page into blocks, and a model classifies each block as main content or filler. The passages relevant to the question are the page’s original text, not summaries written by the model. In an example from Perplexity’s guide, the API extracted relevant passages totalling about 6,800 characters from four pages on HTTP caching. The RFC 9111 standard alone is about 79,000 characters as Markdown. The same guide says two live fetches of the same page 13 seconds apart returned different passages, even though the question and the token limit were the same.
Developers choose how fresh the content must be for each request. By default, the API uses a copy from Perplexity’s index if the copy was fetched within the last two days. Otherwise it fetches the page live. Setting max_age_seconds: 0 always forces a live fetch, and cache_only prevents one. Each result shows whether the content came from a stored copy and when that copy was fetched.
The API fetches URLs in parallel, and each result has its own status. The response carries HTTP status code 200 even if some URLs fail to fetch. The guide therefore says the status of each result must be checked separately.
Perplexity compared the service with Firecrawl, Exa and others on thousands of URLs from 715 websites. A grader model scored the results without knowing which service had produced them. The Context API returned usable content for 92.3 per cent of URLs and recovered 87.9 per cent of the key facts on the pages, the company said. Content precision, the share of returned text that was page content rather than cookie notices, menus or adverts, averaged 80.3 per cent. A failed fetch scored zero. Data Phoenix reported that the published information is not enough to verify the figures independently.
The press release says a single request can hold 10 URLs. The documentation, however, ties the limit to the usage tier: 10, 15 or 25 URLs, or at most 50 with an increase. Failed fetches are not billed. Successful fetches are billable, and so are not_modified results, where the page has not changed since a given time, the documentation says. Scoring pages with yes/no questions is billed separately. Using the API requires a Perplexity API key, and the guides are at docs.perplexity.ai.