API Credits

API credits are the abstract billing unit that most AI search and web extraction vendors meter usage in: you buy or are allotted a pool of credits per month, and each API call debits some number of them. Credits are not comparable between vendors, because one credit can mean one search, one returned result, one page render, one browser-minute, or one LLM extraction pass depending entirely on whose pricing page you are reading.

Firecrawl's published credits table as of July 2026 illustrates the pattern: a scrape costs 1 credit per page, a crawl 1 per page, a map 1 per page, a search 2 per ten results, an interactive browser session 2 per browser minute, and a monitor check 1 per page per check, with a note that advanced features such as JSON output mode cost additional credits on top. Its published plans, billed yearly, run $16 per month for 5,000 credits, $83 for 100,000, $333 for 500,000, and $599 for 1,000,000. Tavily publishes a pay-as-you-go rate of $0.008 per credit with 1,000 credits per month on the free tier. Exa skips the abstraction and prices per request directly: $7 per thousand searches, $1 per thousand pages of contents per content type, $1 per thousand for results returned beyond the tenth, and $12 to $15 per thousand for deep search. Line those up and the incomparability is obvious. A Firecrawl "search" bills two credits for ten results; an Exa "search" bills a flat per-request rate and then charges again for the page text you actually wanted. The word is the same and the billing shape is not.

The normalization that works is price per successful useful result, computed on your own traffic. Take 200 representative queries or URLs from your product, run each candidate vendor at the settings you would actually ship (rendering on or off, result depth, structured extraction on or off, proxy tier), apply your own acceptance check to every response, and divide total spend by the count that passed. Two things break the naive arithmetic. Failures usually still bill: a page that returns a block screen, a CAPTCHA, or empty content generally consumes its credits anyway, so a vendor with a 10 percent lower sticker rate and a 20 percent lower success rate on your sites is more expensive. And depth multipliers move quietly: asking for 25 results instead of 10, or turning on JSON extraction, changes the per-call credit cost without changing the call in your code in any visible way.

Credit pricing exists for a defensible reason. A raw HTTP fetch and a full browser render differ by roughly an order of magnitude in compute, and credits let a vendor charge for that spread without publishing a per-feature price matrix. The cost to the buyer is that headline rates carry almost no information in isolation, and the things that actually determine your bill sit elsewhere in the contract: rate limits, concurrency caps, whether unused credits roll over, whether an annual commitment is required to reach the published per-unit price, and what happens when you exceed the allotment mid-month. Ask for all five before signing.