OpenAI structured outputs from PDFs: the request, the price per page, and the measured fabrication rate

One Chat Completions call, the PDF as a base64 file part, the schema in the prompt text, temperature 0, model pinned to the dated snapshot. 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 every model in the comparison got, "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. The frozen body also carried logprobs: true until the smoke run, when the API declined it with a 403 on a PDF file input. The parameter was removed and the trim recorded in the run manifest, which is why the confidence column for this row reads "none surfaced". The record is in the analysis plan, §12.

The structured variant sends the identical prompt and file and adds a response_format block with the same schema. The model then cannot emit a token that breaks the schema.

The comparison's frozen schemas list every field as required, set additionalProperties to false and keep every leaf nullable, which is how the docs say to express a field that may be absent under strict: true. The same schema bytes went to both rows, so the only difference between them is the response_format block. One corpus schema, cut to three fields:

The docs describe a PDF page as extracted text plus a page image, both sent to the model, and say that raises token usage. As of 2026-09-03 the pricing page has no line on how a PDF page is billed, so the per-page cost below is the measured figure.

Velrim's published comparison of six extraction setups (September 2026) measured ~$1.0 (measured, prompt-cached) per 1,000 pages over 319 pages, with the repeated prompt prefix billed at the cached input rate of $0.075 per million tokens and the answer at $4.50 per million. The Gemini rows in the same table came to ~$2.2 (measured, token-priced).

124 public documents, 2102 fields, 96 of them absent by the audited answer key. Free-decode gpt-5.4-mini invented a value for 10.8% of the absent fields, interval [5.7, 18.4]. With Structured Outputs, 10.4%, interval [5.3, 17.6]. Accuracy moved from 0.735 to 0.718 macro-F1, the highest two rows in the table, against Velrim's 0.691. Both rows completed 372/372 attempts.

The trap rows are the second fabrication measure: fields planted in the schema that cannot exist in that document type, so any answer is an invention. gpt-5.4-mini answered 37.9% [30.1, 46.3] of them free-decode and 39.9% [32.5, 48.4] with Structured Outputs, against 20.7% [15.3, 27.2] for Velrim. On the other side, the model answered 96.1% [94.6, 97.3] of the fields that were on the page and Velrim 88.0% [84.0, 91.2], so Velrim's lower rates are 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.

gpt-5.4-mini with Structured Outputs, when accuracy per dollar is the whole requirement: it took the top two accuracy rows in the comparison at a twentieth of the price. Velrim, when a value needs a number to sort on and a place on the page to check it, and a field that is not on the page must come back empty, which the trap rows measure.