Direct reads matched on a homemade set

The OpenRouter tutorial is signed October 29, 2026. The structured publication time on the page is 00:00:00 UTC that day, an exact hour. The topic board lists the same item at 00:33:30 UTC, which is 08:33 in Beijing and inside this edition's window. The task is stated narrowly. A PDF goes to a model that can read files, together with a strict JSON schema that describes one row. The model is asked to return typed rows. A loss written in parentheses comes back as a negative number, and a table that crosses a page comes back as one list.

The score is from a set the authors built. The page says the five PDFs were generated from known data, so each cell could be checked against that source. Two of them are scans. The list is an invoice, a quarterly report, a 58-row shipment log that runs onto a second page, a scanned purchase order, and a deliberately rough scan with merged cells. Each route ran once over all five PDFs, six pages and 644 cells in total. The comparison is the data used to generate the files. Prices come from the model catalog as of October 8, 2026. The page says these results describe this test set, not model performance in general.

Every model that read the PDF directly got all 644 cells right, the page says. The failures came from the route the PDF took to reach the model, and two of those failures returned a normal response with no error. A scan sent through the free cloudflare-ai parser returns zero rows and no error. A later sentence states the same split another way: the free parser got all three digital PDFs right and both scans completely wrong. When a model cannot read PDFs, scans need the mistral-ocr parser. Long tables can stop at the output token limit, and the page says to reject a response whose finish reason is length.

The cost comparison uses that same six-page run. GPT-6 Luna and Claude Haiku 5.5 were the cheapest routes that got every cell right, including the rough scan, and both list at 0.10 dollars per million input tokens. Gemini 3.8 Flash was just as accurate and cost eight to ten times as much per page on the same files. Its list price is 1.50 dollars per million input tokens, and the page separately says that price was discounted to 0.75 dollars on October 21, 2026. The OCR route adds 2 dollars per 1,000 pages on top of the token price. The two misses the page names are footnote markers on Enterprise and Hardware resale: the schema asked for them to be removed, and the OCR route kept them. The cost per 1,000 pages is the six-page bill scaled up.

The same files also went through pdfplumber. It found every table in the three digital PDFs and returned each cell as a string, such as an amount with thousands separators or a loss in parentheses. The quarterly report's two-level header came back as two rows, with empty values in the merged cells. The shipment log came back as two tables, one per page. It found nothing in either scan, because a scan has no text layer. The page then splits the work: a library is enough when the PDFs are digital and each table has a single header row; once merged headers, page breaks, and number formats need cleanup, the schema is less work to maintain. That last clause is the authors' judgment about maintenance.

What the page leaves out limits how far the 644 can travel. Each route ran once. There is no second run. The catalog date is October 8, the Flash discount date is October 21, and the tutorial date is October 29. Those three dates are not one price list. A perfect cell count belongs to models that read these five generated PDFs directly. It is not an external board. The picture with this column is the diagram at the top of the tutorial: a table feeding rows. The cell count is in the body, not on the drawing.