What the Minexa API actually costs to run at scale (and how it compares to LLM extraction)

Most developers who start evaluating extraction tools focus on accuracy first. Cost comes second, usually after the first invoice arrives.

This post is about that second conversation. Specifically, what it actually costs to run a data extraction pipeline at meaningful volume, and how the Minexa API compares to LLM-based extraction across different page volumes and HTML formats.

The numbers here are not estimates. They come from real page size measurements across six content types: job listings, ecommerce product pages, property results, review pages, search results, and hotel booking pages.

The HTML size problem nobody talks about upfront

When you pass a webpage to an LLM for extraction, you have two options. You can strip the HTML first, removing scripts, styles, and attributes, keeping only the DOM-rendered tags and text. Or you can pass the full HTML as-is.

Stripped HTML averages around 38,965 tokens per page. Full HTML averages around 572,739 tokens per page. That is roughly a 15x difference in token volume per page.

This matters because LLM pricing is token-based. Every page you process costs proportionally more as token count increases. Minexa's pricing is credit-based and flat per page. Whether a page is 10,000 tokens or 800,000 tokens, the credit cost does not change.

Stripping HTML sounds like a reasonable solution until you consider the tradeoffs. Some data you need lives in attributes, not visible text. Stripping aggressively can silently remove the exact field you are trying to extract. And if you implement a token cap instead, you risk truncating the page mid-content with no error signal. There is no clean answer on the LLM side.

What the numbers look like at three volume tiers

The comparison below covers three monthly volumes that correspond roughly to Minexa's three plan tiers: 10,000 pages, 120,000 pages, and 2,000,000 pages.

At 10,000 pages per month (stripped HTML): The cheapest available LLM, GPT-5 nano, costs approximately $24. Minexa Personal is $15 as a flat monthly floor and includes 10,000 credits. Mid-range models like GPT-5 mini cost $117 at this volume, and Claude Haiku 4.5 reaches $440. Even at the smallest scale, the cost gap is already visible.

At 120,000 pages per month (stripped HTML): GPT-5 nano costs $285. Minexa Startup handles the same volume for $60. GPT-4o-mini costs $773, GPT-5 mini costs $1,410, and Claude Haiku 4.5 costs $5,280. Only the cheapest nano-class models stay below $1,000 for this volume. Minexa Startup handles it for $60.

At 2,000,000 pages per month (stripped HTML): GPT-5 nano costs $4,700. GPT-4o-mini costs $12,900. GPT-5 costs $117,400. Claude Sonnet 4.6 reaches $263,700. Minexa Business covers the same volume for $500.

Switch to full HTML and every LLM figure above multiplies by roughly 15. At 10,000 full pages, GPT-5 nano costs $290, GPT-4o-mini costs $860, and Claude Sonnet 4.6 costs $17,332. Minexa Personal remains at $15. At 120,000 full pages, even GPT-5 nano costs $3,480, which is 58 times more than Minexa Startup at $60. At 2,000,000 full pages, GPT-5 nano costs $58,000 versus Minexa Business at $500.

Minexa's cost is identical whether you process stripped or full HTML. There is no token-based pricing. Page size is irrelevant to the credit calculation.

The indirect costs that do not appear in token pricing

Token cost is only part of the picture. LLM extraction pipelines carry additional overhead that compounds at scale.

Field naming is inconsistent across responses. One page returns "salary_range", another returns "compensation", another returns "pay". Downstream normalization logic is required to unify these into a consistent schema before the data is usable. That normalization step takes engineering time to build and maintain.

Retry logic is another hidden cost. LLMs processing complex or noisy HTML occasionally return malformed JSON, truncated output, or values assigned to the wrong field. A production pipeline needs retry handling, validation checks, and alerting for these cases. At 100,000 pages per month, even a 1% error rate means 1,000 rows requiring review or reprocessing.

Minexa returns the same column names on every page processed with the same scraper. The scraper is trained once via the Chrome extension, which generates a stable scraper_id. Every subsequent API call using that scraper_id returns identically structured JSON. No normalization step is needed. If a value is missing from the page, the field returns null rather than a fabricated default.

To run a batch extraction via the Minexa API, the POST request to https://api.minexa.ai/data/ includes the scraper_id, a columns parameter specifying which fields to return (using "top_30" for automatic selection or explicit field names), the list of urls to process, scraping configuration, and a threads value controlling parallelism. Up to 50,000 URLs can be submitted in a single batch request.

The ready-to-use Python script available from the extension's API Request button handles pagination across the response automatically, writes checkpoint files after each iteration as JSON, CSV, and Excel, and continues until the full job is complete. This means a partial run is never lost if the script is interrupted.

If you want to explore the full API reference, the documentation is at minexa.stoplight.io/docs/minexa.

Three real switching scenarios

A retail analytics team was tracking competitor prices across roughly 80,000 product pages per month using a GPT-5 nano pipeline on stripped HTML, including approximately 20% retry overhead. Monthly extraction cost was around $230. After training a single Minexa scraper on the product page structure, the same workload ran on the Startup plan at $60 per month. Price fields were always pulled from the correct DOM element, and no prompt updates were needed when the retailer updated its page layout.

A property data firm was feeding full HTML from listing pages into a language model to extract price, square footage, location, and listing date. At 200,000 pages per month, token costs exceeded $29,020 per month using GPT-5 mini on full HTML. Switching to Minexa with a trained scraper per listing type reduced the monthly extraction cost to $500 on the Business plan. The LLM had been occasionally swapping asking price and last sale price. Minexa eliminated the issue entirely by binding each column to its correct DOM element.

A sales team was extracting company name, contact details, industry, and employee count from roughly 50,000 directory pages per month using Mistral Small 2 on stripped HTML. Inconsistent JSON field naming across responses required downstream normalization logic. After switching to Minexa, the same columns were returned with consistent names and structure on every page, the normalization step was removed, and the monthly extraction cost dropped from approximately $485 to $60.

Latency at scale

These cost figures assume HTML has already been fetched. Upstream scraping time is excluded from both sides of the comparison since both approaches face the same rendering cost.

For the extraction step itself, Minexa processes a page in hundreds of milliseconds to a few seconds with parallel thread support across the plan's thread limit. LLMs processing hundreds of thousands of tokens per page are significantly slower per page and do not benefit from the same parallelism. At hundreds of thousands of pages per month, this difference translates to hours or days of additional processing time.

Where LLMs still make sense

Below roughly 10,000 pages per month on stripped HTML, the cheapest nano-class models (GPT-5 nano, GPT-4.1 nano, Gemini Flash Lite) come close to Minexa Personal in cost. If volume is very low and the extraction schema changes frequently, an LLM pipeline can be quicker to adjust without retraining. That tradeoff is real and worth acknowledging.

Beyond that threshold, and especially on full HTML at any volume, the cost and consistency advantages of DOM-based extraction become substantial and widen as volume increases.

If your pipeline is already at production scale or approaching it, the Minexa API documentation and plan details are at minexa.ai/plans. The Chrome extension for training scrapers is available at the Chrome Web Store. Most developers have a working scraper and their first API extraction running within a single session.