Say you're using an AI research tool built on OpenAI's Responses API to examine a pitch deck. You want to know what the plotted evidence shows, not just how the founders describe it.
For that task, the documented input route is a PDF sent to a vision-capable model. OpenAI sends both extracted text and images of the pages, so the chart and its accompanying explanation can reach the model together.
Sending the PowerPoint file directly gives it the text, but not the embedded charts or images. That's existing file-input behavior, not a judgment by the model about which parts matter.
A summary of the written argument may have no need for those page images. Examining the chart does, and including them can increase token usage.
Source: https://developers.openai.com/api/docs/guides/file-inputs
For that task, the documented input route is a PDF sent to a vision-capable model. OpenAI sends both extracted text and images of the pages, so the chart and its accompanying explanation can reach the model together.
Sending the PowerPoint file directly gives it the text, but not the embedded charts or images. That's existing file-input behavior, not a judgment by the model about which parts matter.
A summary of the written argument may have no need for those page images. Examining the chart does, and including them can increase token usage.
Source: https://developers.openai.com/api/docs/guides/file-inputs