AI scraping API pricing

What an AI extraction API is

An AI extraction API turns web content or a browser state into a requested schema, answer, table, markdown representation, or agent-ready object. The defining job is transformation into useful structure, not merely downloading the source page. Some products accept a URL and perform retrieval as part of the call, while others operate on content, tasks, robots, workflows, or browser sessions. Firecrawl documents separate extraction approaches and token-sensitive extraction, while Diffbot describes extraction and crawl operations under credits. (Firecrawl extraction guide, retrieved 2026-07-26; Diffbot pricing, retrieved 2026-07-26.)

The closest neighbour is a scraping API. A scraping API’s primary contract is page retrieval: give it a URL and receive fetched or rendered content. AI extraction adds semantic transformation into fields or answers. A product that performs both jobs may expose one combined credit weight, two separate meters, or a plan allowance that leaves the combination unclear. ScrapingBee documents extraction alongside its scraping plans, and Zyte API publishes page retrieval plus extraction and token add-ons. (ScrapingBee pricing, retrieved 2026-07-26; Zyte API pricing, retrieved 2026-07-26.)

A cloud browser provides a controllable session, while AI extraction interprets what the session or page contains. A web search API finds relevant resources, while extraction turns located resources into a schema or answer. Agent products can combine browsing, search, and extraction in one workflow, but that does not make browser time, search calls, model tokens, and completed tasks the same billing unit. Browserbase publishes separate browser, Search, Fetch, and Extract allowances, and Browser Use combines browser, proxy, task, step, and model consumption. (Browserbase pricing, retrieved 2026-07-25; Browser Use pricing, retrieved 2026-07-26.)

What this category bills on

Page and request meters are the simplest form. A vendor can charge one extraction request or apply an extraction credit multiplier to each fetched page. Some products charge per completed task, robot action, captured row, screenshot, crawl page, or monitoring run. Browse AI, for example, defines robot-task consumption through captured rows, screenshots, detail pages, and premium-site rules rather than a universal page request. (Browse AI pricing, retrieved 2026-07-26; ScrapeGraphAI, retrieved 2026-07-26.)

Token meters make input size, output size, schema complexity, and model use part of the bill. Browser-time and compute meters make workflow duration or resources relevant. Hybrid products can charge retrieval, proxy traffic, browser time, and extraction separately, so a published extraction unit is not automatically an all-in rendered-page unit. Firecrawl documents token-sensitive extraction and separate browser-minute consumption, while Jina records Reader output tokens and Search request tokens. (Firecrawl pricing, retrieved 2026-07-26; Firecrawl browser documentation, retrieved 2026-07-26; Jina AI Reader, retrieved 2026-07-25.)

The comparison workload

The table combines one JavaScript-rendered page retrieval with one AI extraction operation and shows every product carrying ai-extraction in the dataset. A row is assumption-free only when the public record supplies the meters required for both parts. If one modifier is absent, the model provisionally treats it as included and marks the row as assumption-dependent. Products sold by task, compute, browser time, token, workflow, or an unpublished contract remain visible without an invented page billing unit. (Scraping API Prices methodology, retrieved 2026-07-26.)

Rows are ordered by evidence state before bill size: assumption-free calculations, assumption-dependent calculations, published lower bounds, quote-required rows, and rows whose public meter does not fit the workload. Affiliate data does not affect inclusion or order. An assumption-marked row is a question to resolve with the vendor, not evidence for a cheapest claim. (Scraping API Prices methodology, retrieved 2026-07-26.)

What can change the bill

Schema size, input and output tokens, model selection, crawl depth, browser duration, proxy route, retries, screenshots, monitoring cadence, and human review can change total consumption or add a second meter. A workflow that extracts from several pages should count retrieval and extraction for every page unless the contract explicitly defines a different bundle. A task-based product should be budgeted from the vendor’s task definition instead of being forced into a page comparison. (Firecrawl extraction guide, retrieved 2026-07-26; Anchor Browser, retrieved 2026-07-25; Browse AI pricing, retrieved 2026-07-26.)

Products whose public meter does not produce this table’s bill

• ScrapeOps combines proxy, rendering, anti-bot, target, and AI extraction credits. The comparison workload reaches a custom offer, and the pricing page is the route to a contract for credits, concurrency, and request rules. Retrieved 2026-07-26.
• Zyte API provides HTTP retrieval, browser-rendered retrieval, extraction, and token add-ons. The applicable comparison state requires a quote, and the pricing page is the route to workload-specific volume terms. Retrieved 2026-07-26.
• AgentQL combines per-call structured extraction with remote-browser time. The record does not publish the basic rendered-page billing unit needed by this workload; the product page is the route to the entry tier or a managed dataset and deployment contract. Retrieved 2026-07-25.
• Anchor Browser offers completed-task credits plus infrastructure meters for browser creation, active time, proxy traffic, and AI steps. Those units do not yield a generic rendered-and-extracted page bill; the product page is the route to the entry tier or a custom concurrency contract. Retrieved 2026-07-25.
• Browser Use Cloud combines browser time, proxy traffic, task initialization, model tokens or agent steps, and optional skills. It has no generic rendered-page billing unit for this workload; the pricing page is the route to the entry tier or custom limits and deployment terms. Retrieved 2026-07-26.
• Cloudflare Browser Rendering provides browser automation through Workers with browser-time and concurrency meters. It does not publish a generic AI extraction request billing unit, and its public documentation is the purchase path. Retrieved 2026-07-25.
• Diffbot provides page extraction, crawling, proxy extraction, natural-language processing, and Knowledge Graph operations under credits. Its public units do not yield the combined request used here; the pricing page is the route to the entry tier or a custom credit and throughput contract. Retrieved 2026-07-26.
• Jina AI Reader and Search Foundation APIs provides Reader and Search products under token consumption. The public widget did not expose current package pricing or a generic rendered-page meter; Jina’s sales contact is the contract path. Retrieved 2026-07-25.
• Kadoa provides consumption-based data extraction and monitoring workflows, but its public pricing page publishes plan descriptions without a billing unit or rate. The same page is the route to the entry tier or enterprise usage and service terms. Retrieved 2026-07-25.
• ScrapeGraphAI combines scraping formats, structured extraction, search, crawling, monitoring, and optional stealth under shared credits. Its public meters do not define the generic rendered-page request used here; the product page is the route to the entry tier or custom credits and limits. Retrieved 2026-07-26.
• Spider.cloud provides crawling and extraction-capable workflows through bandwidth-and-compute or concurrency-based offers. Neither contract converts to the combined page workload without assumptions; the pricing page is the route to a public plan or custom arrangement. Retrieved 2026-07-26.

How to use the comparison

Require written confirmation of both the retrieval meter and the extraction meter. Define the schema, expected content size, model path, rendering need, proxy route, and number of pages before comparing contracts. Add token, task, browser, compute, bandwidth, or monitoring meters when the selected product exposes them. Treat assumption-marked and not-computable rows as procurement questions, not as evidence that a competitor is cheaper or absent. (Zyte API pricing, retrieved 2026-07-26; Firecrawl extraction guide, retrieved 2026-07-26; Scraping API Prices methodology, retrieved 2026-07-26.)