Gemini 2.5 Flash PDF extraction: the request, the price per page, and the measured fabrication rate

One generateContent call, the schema in the prompt text, the PDF as a base64 part, temperature 0 and nothing else set. This is the body Velrim's published comparison of six extraction setups (September 2026) sent for every document, with a two-field schema in place of the corpus schemas.

The prompt text is the sentence "Extract the following fields from the document. Return JSON matching this schema. Use null for fields not present in the document." followed by the JSON Schema. Every model in the comparison got these bytes. The same body runs on the Vertex AI endpoint for the same model pin, https://aiplatform.googleapis.com/v1/projects/<project>/locations/global/publishers/google/models/gemini-2.5-flash:generateContent, which is the route the comparison used.

The constrained variant sends the identical prompt and adds a MIME type and the schema to the generation config. The model then cannot emit a token that breaks the schema.

As of 2026-09-03 the reference marks responseSchema deprecated and keeps responseJsonSchema with the list of supported keywords above. Google's structured outputs guide now documents the Interactions API on Gemini 3 models and does not mention gemini-2.5-flash or generateContent.

The model card lists text, images, video and audio as inputs, a 1,048,576-token input limit, a 65,536-token output limit and a January 2025 knowledge cutoff (Gemini 2.5 Flash model card, read 2026-09-03).

At 258 tokens a page and $0.30 per million input tokens, the pages themselves cost $0.08 per 1,000. The measured figure in Velrim's published comparison of six extraction setups (September 2026) came to ~$2.2 (measured, token-priced) over 319 pages at the paid tier. The gap is prompt and schema tokens on the input side and answer tokens at $2.50 per million on the output side. The free tier is priced at zero and, per the pricing page, the inputs are used to improve Google's products.

124 public documents, 2102 fields, 96 of them absent by the audited answer key. Free-decode Gemini invented a value for 17.0% of the absent fields, interval [10.1, 26.7]. With constrained decoding, 12.8%, interval [7.6, 20.5]. Accuracy moved from 0.710 to 0.697 macro-F1. Pooled across all six setups the rate is 17%. Velrim, on the same model: 11.5%.

The two Gemini rows completed 328/372 and 315/372 attempts. The shortfall was rate limiting on a new Google Cloud project at its default quota, plus one document that twice came back without usable JSON. Google's status page was clean for the window. On the other side of the fabrication column, Gemini answered 96.5% [95.0, 97.7] of the fields that were in the document and Velrim 88.0% [84.0, 91.2], so the lower invention rate is paid for in blanks.

The score is one number per field, 0 to 1, with its error curve on the error curves page and the definition of the fabrication measure on the metric page. Grounding is a page number, a box and the matched text from the PDF's own word map, no second model call. Price is $0.02 per extracted page, prepaid from $5, about 500 pages free with no card.

Gemini direct, when the schema is small, every field is always on the page and a wrong value costs little, at a tenth of the price. Velrim, when a value needs a number to sort on and a place on the page to check it, and a blank is cheaper than an invention.