ChatGPT + API silently drops part of scanned PDF pages | OpenAI | Page 1

Reported by @austere furnace

Bug Report: ChatGPT + API silently drops part of scanned PDF pages

`Steps to Reproduce`

Take any PDF with /Rotate 270 and a landscape-stored bitmap (standard output of Brother/HP/Canon/Xerox MFPs and fax-to-PDF
gateways).
Send via /v1/responses with input_file, or upload to ChatGPT web.
Ask about content near the top of the displayed page.

`Expected Result`

Model reads the content as it appears in any compliant viewer. /Rotate is ISO 32000-1 §7.7.3.3, part of PDF since 1.0 (1993); all
standard viewers honor it.

`Actual Result`

Model returns "I can't determine..." or hallucinates. The top of the displayed page is silently absent from the model's input.

gpt-5.4, gpt-5.3-codex, gpt-4.1-mini: fail 100%
gemini-2.0-flash on the exact same files: reads correctly 100%
Reproduced at 5 resolutions from 8.7M down to 87K pixels, so it's not a tiling/downsampling artifact.

`Environment`

Responses API (/v1/responses with input_file) AND ChatGPT web. gpt-5.4, gpt-5.3-codex, gpt-4.1-mini

Additional Information

Please provide relevant details to help resolve the issue, such as:

ChatGPT Shared Link (if applicable).
Screenshots or videos demonstrating the problem.

-# ➜ Need to contact support? Visit the OpenAI Help Center.

#ChatGPT + API silently drops part of scanned PDF pages