Voice mode loses accurate PDF page context after switching from web chat

What issue are you seeing?

When I provide a PDF URL in the normal ChatGPT web interface, the model can correctly locate and discuss specific pages in the PDF. However, if I then switch to Voice Mode in the same conversation and ask it to continue teaching/explaining the PDF, it can no longer reliably locate the correct page and starts referring to the wrong slide/page content.

Example PDF:
https://cs231n.stanford.edu/slides/2025/lecture_5.pdf

In normal text/web chat, I can ask it to start from a specific page (for example, page 80) and it can identify the corresponding slide content accurately. After switching to Voice Mode, asking it to continue from a specific page often results in it describing content from a different page or otherwise losing the document position/context.

This makes Voice Mode difficult to use for interactive teaching over PDFs, because the conversation context appears to carry over but the document/page grounding does not remain accurate.

Steps to reproduce

1. Open ChatGPT on the web.
2. Send this PDF URL: https://cs231n.stanford.edu/slides/2025/lecture_5.pdf
3. In normal text mode, ask the model to locate and explain a specific page, e.g. page 80.
4. Confirm that it is discussing the correct slide/page.
5. Switch to Voice Mode in the same conversation.
6. Ask it to continue teaching from that page or move to a nearby specific page (e.g. page 81, 101, 104).
7. Observe that Voice Mode may identify the wrong slide/page or describe content that does not correspond to the requested page.

Expected behavior

Voice Mode should preserve the same PDF/document grounding available in the normal web conversation. If the text-mode model can locate page N accurately, switching to Voice Mode in the same conversation should allow it to continue referring to page N and subsequent pages accurately.

Actual behavior

After switching to Voice Mode, the model appears to lose or degrade its ability to resolve the PDF's page-level context, even though the conversation itself is preserved.

Why this matters

A major use case for Voice Mode is tutoring over lecture slides, papers, and textbooks. Accurate page-level continuity is essential: users should be able to say things like “continue with page 101” or “go back to page 80” without the model silently teaching from a different slide.

Environment

• ChatGPT web interface
• Voice Mode entered from an existing text conversation
• PDF supplied via public URL rather than uploaded directly
• Reproducible with Stanford CS231n Lecture 5 (2025) PDF above

I realize this repository is primarily for Codex, but I'm reporting it here because this appears related to context/tool parity across ChatGPT interaction modes and there may not be a more appropriate public issue tracker.

The report names the ChatGPT web interface, Voice Mode, and the Stanford PDF, but no Codex file, test, or entry point. Start by checking whether this repository owns the reported behavior and reproduce the text-to-Voice transition with the supplied URL; done would mean page references remain accurate after switching modes.