PDF file input in Chat Completions works with some API keys but fails with “No content could be read” for others

API Bugs
1

Hi OpenAI team/community,

We’re seeing inconsistent behavior when sending PDF files to an OpenAI model through the Chat Completions API.

The same request, same PDF, and same model can successfully read and answer questions about the PDF when using one API key, but fail to read the document when using another API key.

Example request

payload = {
    "messages": [
        {
            "content": [
                {
                    "text": "what questions are in this document?",
                    "type": "text"
                },
                {
                    "file": {
                        "file_data": f"data:application/pdf;base64,{pdf_b64}",
                        "filename": "sample.pdf"
                    },
                    "type": "file"
                }
            ],
            "role": "user"
        }
    ],
    "model": "gpt-5.6-luna"
}

Behavior with API key A

The request succeeds and the model correctly answers based on the contents of sample.pdf.

Behavior with API key B

The API request itself completes successfully (finish_reason='stop'), but the model responds that it could not access the PDF:

I can help, but I can’t currently extract text/questions from your file because
the PDF content didn’t load (“No content (images or text) could be read from
document 'sample.pdf'”).

The important point is that this is reproducible with the same PDF and same request payload. Switching the API key changes the result.

We have therefore ruled out the obvious possibility of the PDF simply being unreadable, since the document is parsed correctly when the request is made using another API key.

Questions

Is PDF/file-input support subject to any of the following at the API key, project, organization, or account level?

  • Feature rollout / allowlisting
  • Different backend routing
  • Project or organization configuration
  • API tier or account eligibility
  • File-processing capability flags
  • Regional differences
  • Model entitlement differences

Is there a way to determine from the API whether a particular project/API key supports PDF document processing?

Also, when this happens, is there any request ID, response metadata, or diagnostic information we should collect to help OpenAI Support identify why the document-processing stage failed?

The most confusing aspect is that the request does not return an API error. It returns HTTP success and a normal assistant response, but the model apparently receives an internal indication similar to:

No content (images or text) could be read from document 'sample.pdf'

Any guidance on what could cause this behavior to differ between API keys/projects would be appreciated.

Thanks!

3 Likes