Skip to content
MakerGain
← Back to cases

A five-credit extraction that tells you one is enough when you only need the body text

Edited by MakerGainPublished: Article updated:

ScrapeGraphAI's Extract turns a web address, HTML, or Markdown plus a sentence of instruction into JSON. It charges 5 credits per call; enabling stealth adds 5 more; failed or error-returning requests are not charged; PDF costs 1 credit per processed page. The docs suggest switching to scrape, which costs 1 credit, when all you need is readable Markdown, and note to cache web addresses that do not change often.

Imagine an utterly ordinary task: every morning, copy the model numbers, prices, and stock availability from dozens of product pages into a spreadsheet. Writing the first version of a scraper is not hard; the hard part is that when a page is redesigned, the XPath and CSS selectors fail together, the program does not throw an error, and it just quietly returns empty values. ScrapeGraphAI is aimed at exactly this step—you give it a web address and a sentence in plain language, and it gives you back a piece of JSON.

Its Extract service takes the input as one of three choices: a web address, raw HTML, or Markdown; then you give it a sentence in natural language describing what to extract, and a schema is optional, used to constrain the shape of the output JSON. The official docs state the pitch plainly: replace fragile selectors such as XPath with AI context extraction, accept any URL, raw HTML, or Markdown, and support structured results output according to a JSON Schema. [1]

Image: source ScrapeGraphAI[1], retrieved 2026-11-14, original interface preserved.

The page-fetching part also has adjustable settings. In the same documentation, page fetching can use JS rendering mode, include request headers, and enable stealth mode when necessary. [1]

Billing is the most interesting part of this, but first you have to separate which calls are charged and which are free. Extract charges 5 credits for each successful call; enabling stealth in the page-fetching settings adds another 5 credits; choosing auto, fast, or js for the rendering mode does not affect the cost. The free scope stated in the docs is that “failed or error-returning requests are not charged.” The harder the page is to handle, enabling stealth instead adds 5 more credits per order. So in this price list, “hard” and “failed” are two different things: difficulty carries an explicit surcharge, and only failure is free.

This distinction is not wordplay. Firecrawl's pricing page draws finer lines: a scrape that returns an empty result is not charged, but error status pages such as 403/404 are still returned to you and cost 1 credit. [2]

PDF is also counted separately: 1 credit per actual processed page. [1]

More intriguing is that it teaches people to spend less. The credits documentation suggests: when you only need readable page content and do not need structured extraction, use scrape and specify MarkdownFormatConfig to output readable Markdown (1 credit), rather than extract (5 credits); cache the URL that does not change often to reduce repeated fetching. [3]

The documentation lays out the difference between the two deliverables: extract returns JSON organized according to the instructions, while scrape, after MarkdownFormatConfig is specified, converts page content into readable Markdown; the cost composition of the price difference is not explained.

Subscription tiers are layered by credits volume, throughput, and support. Free is $0/month with 500 API credits per month; paid tiers start at Starter, at $17/month with 10,000 credits. The docs say every plan can access every service, and higher tiers add credits, throughput, and support. [4]

Putting “500 credits” and “5 credits per Extract” together gives 100 ordinary Extracts; enabling stealth will use them up faster, while PDF is billed at 1 credit per actual processed page. This is a conversion based on the unit price of an ordinary call, not per web page; the docs do not describe this conversion.

It is not the only player. Firecrawl's pricing page (effective 2026-09-04, USD) says: basic scrape/crawl costs 1 credit per page; using the JSON, Question, or Highlight formats adds 4 credits per page; Hobby is $19/month (equivalent to $16/month with annual billing) and includes 5,000 credits; Standard is $99/month (equivalent to $83/month with annual billing) and includes 100,000 credits. [2]

The two companies made a similar choice on the same condition: structured extraction carries a surcharge. At least on these two price lists, structured delivery is priced separately; the pages do not give the cost composition. There are also real differences—the rules on whether failed and error pages are charged are simply different.

Founder Lorenzo Padoan's personal homepage bio says he took ScrapeGraph-ai to 22,000+ GitHub stars and also says he made an open-source project profitable; the latter does not name the project, neither item has a date, there are no profit figures, and there is no explanation for why this credits price list is set this way. [5]

Open-source stars at most show that people are paying attention; payment for the hosted API is what shows that people are opening their wallets, and the latter has no public numbers. Current revenue, customer count, and cost structure are all undisclosed, and that profitability claim cannot serve as evidence of the hosted business's profit today; whether this business makes money is not something this article can judge. What can be confirmed is that the documentation is aimed at developers writing the instructions and the schema, and that the paying party is subscription accounts; this does not amount to verified active paying users.

This price list separates “hard” from “failed” and also prices structured delivery separately; what is most memorable is not which tier is cheaper, but how much more customers are willing to pay to turn the same page content into JSON. The public documentation gives the price, not the answer.

Continue reading

AI developmentAI productivity
AI2SQL: AI Writes the SQL; You Press Run on the Read-Only Connection

AI2SQL turns natural language into SQL based on real table structures and executes it on a read-only connection; founder Mustafa Ergisi chose to bootstrap, grew by contacting free users one by one and through SEO and word of mouth, and admits he learned costly lessons on retention and pricing. The product sells on the read-only boundaries of sqlGuard and MCP, with three tiers priced at $9, $19, and $39, but the TrustMRR page fields and the founder’s message give two sets of MRR.

AI development
When AI Gets the Refund Rule Wrong, Langfuse Wants You to See What It Checked

Langfuse records one LLM application interaction as an expandable trace, letting teams review retrieved content, prompt versions, parallel steps, and token/cost, then ties evaluation back to datasets and prompt experiments. The team moved from Finto to open-source LLM developer tools; early customers pushed it to add prompt management, and it was ultimately acquired by ClickHouse in January 2026.

AI developmentAI searchAI productivity
Customers Can’t Use the Product, Bad Reviews Follow: PDF.ai Turns Manuals Into an Interface Billed by the Page

While at Cisco, Damon used 6 months of unpaid leave to build side projects; the first four had zero revenue, and the fifth, Testimonial, began generating revenue. E-commerce customers being unable to use a product and leaving negative reviews led him to explore letting customers talk to PDF manuals, which became the Parse, Extract, and Ask API from PDF.ai: billed by page, caching docId avoids repeated parsing, but extraction is still billed by page and asking by total pages.

AI gamesAI developmentAI images
About Ten Prototypes, Only One Kept: A 25-Year Game Industry Veteran Takes a Game to Steam Solo with Rosebud

Damian, a 25-year game-industry veteran, used Rosebud to make about ten prototypes. After family and friends playtested them, he kept only Hollow Signal and, on July 14, took it to Steam solo under the name Dev Solo Studio. The roguelike shooter moved from auto-aim to directional shooting and added eight languages, controller support, and accessibility support, showing the trade-offs of taking a prototype through to a full release.