If ChatGPT starts uploading your PDF and then shows “Unknown Error Occurred,” stop resubmitting the same file. PDF is a supported format, so this message does not tell you whether the document, your upload allowance, the browser or network, or ChatGPT itself failed.
Run two controls before changing the original: upload a tiny known-good PDF, then try the original in a fresh official client or private browser session. Change one variable at a time and record each result.
| Known-good PDF | Original in fresh client | What the result means | Next move |
|---|---|---|---|
| Works | Fails | The problem is probably file-specific. | Preserve the original and normalize a copy. |
| Fails | Works | The current client or network is the likely cause. | Isolate extensions, VPN, proxy, DNS, device, and network. |
| Fails | Fails | Look at quota, account state, or a service incident. | Stop editing PDFs; check limits and live status. |
| Works | Works | The failure was transient or session-specific. | Continue in the working session and do not “repair” the original. |
Retry once only after the controls identify a branch. If every controlled upload still fails across two official clients, collect the timestamp, timezone, platform, screenshot, and request ID if shown, then contact support.
Diagnose the Error in the First 60 Seconds
The two controls are useful because they answer different questions. A known-good PDF tests whether ChatGPT can accept any ordinary PDF in the current environment. A fresh client tests whether the original file can upload without the current session, site data, extension set, or network path interfering.
Your control PDF should be simple: one or two pages, selectable text, no password, no form fields, and no embedded attachments. It is a diagnostic control, not a new unofficial size limit. The current OpenAI supported file types page explicitly includes PDF.
For the environment control, use chatgpt.com in a private window or a first-party ChatGPT app. Sign in to the same account if possible. Do not simultaneously compress the file, switch accounts, disable extensions, and change networks; if the upload starts working, you will not know which variable mattered.
The progress bar is not a diagnosis. An upload that fails at 40%, 50%, or near completion may have transferred bytes before a later validation, parsing, quota, or service step failed. Use the controlled outcomes, not the apparent percentage, to choose the next branch.
If Only One PDF Fails, Normalize a Copy
When the known-good PDF uploads but the original still fails in a fresh client, stop resetting ChatGPT. The evidence now points toward the document.
Preserve the original first. Then inspect a duplicate for features that make the file a useful suspect:
- a password, encryption, or restricted permissions;
- interactive forms, signatures, layers, portfolios, or embedded attachments;
- a scan with no selectable text;
- a file that opens with warnings or renders differently in two PDF viewers;
- an export produced by an unusual driver or a partially completed download.
These are diagnostic hypotheses, not an official cause list for the literal error. OpenAI does not publish a cause table for “Unknown Error Occurred,” so the result of the control test matters more than guessing which internal PDF object is wrong.
The safest normalization test is to create a new PDF copy from a viewer or the source document. On macOS, open the duplicate and use Print → Save as PDF. On Windows, print the duplicate to Microsoft Print to PDF. If you still have the source document, exporting a fresh PDF from that source is better than repeatedly rewriting the only copy.
Printing or exporting can flatten interactive elements and rebuild the PDF structure, which is why community reports sometimes find that it works. It is not guaranteed, and it can remove forms, links, signatures, layers, accessibility tags, or metadata. Compare the new copy with the original before uploading anything sensitive or legally important.
Verification: upload the normalized copy once in the same fresh client where the original failed. If it works, keep the original as the record copy and use the normalized version only for ChatGPT. If it fails while the known-good control still works, stop generating more variants; extract only the needed pages or text, or escalate with a sanitized sample that reproduces the problem.
If ChatGPT instead says the file type is unsupported or asks you to try a PDF, that is a different route. Use the file type not supported guide rather than treating a literal format rejection as this generic-error branch.
If Every PDF Fails, Check the Correct Limit Layer
If the known-good PDF also fails, editing the original is wasted work. Check account and upload boundaries before changing more files.
As verified on July 12, 2026, OpenAI's File Uploads FAQ states a 512 MB hard limit per file and a 2-million-token cap for text and document files. Those are separate constraints: a text-heavy PDF can exceed the token cap while staying far below 512 MB.
Upload-rate limits, Project capacity, and shared storage are separate again. A file may be valid by size and still fail because the account has exhausted a rolling upload allowance, a Project is full, or shared storage is at capacity. OpenAI also warns that failed upload attempts can sometimes count toward the rolling cap. That is the practical reason to stop a blind retry loop.
Use this sequence:
- Confirm the PDF is below the documented per-file boundary.
- Try uploading outside the current Project or GPT, if the workflow allows it.
- Check whether recent uploads or failed attempts may have consumed the rolling allowance.
- Remove genuinely unneeded stored files only if storage or Project capacity is the suspected layer.
- Wait for the applicable rate window instead of splitting one PDF into many uploads.
Do not infer an unofficial 10 MB limit from a generic error. For the plan, Project, rate, storage, and file-cap details—and any current official conflicts between help pages—use the dedicated ChatGPT file upload limits guide.
If the Same File Works Elsewhere, Isolate the Client or Network
When the original uploads in a private window or another official client, the PDF is not your first repair target. Return to the failing environment and test reversible changes one at a time.
OpenAI's current ChatGPT error troubleshooting guide recommends this general isolation path:
- Start a new chat and reload the page.
- Sign out and back in.
- Clear ChatGPT site data or cookies, then retest.
- Disable browser extensions, especially privacy, script-blocking, download, or security tools.
- Turn off a VPN, proxy, Web Protect feature, or secure-DNS filter temporarily.
- Try another supported browser or first-party app.
- Try another device or network.
A private window is a fast control because it usually starts with a clean extension and site-data state, but it does not isolate everything. The same device may still use the same VPN, DNS filter, endpoint security software, and network. If private mode works, first compare extensions and site data. If it fails too, compare the network path next.
Verification: after each change, retry the known-good PDF once. When it works, reintroduce the disabled layer and test again if doing so is safe. That confirmation separates a real cause from a transient success. Stop when you identify the smallest stable working path; a full browser reset should not be the opening move.
If Both Controls Still Fail, Check Service Status and Escalate
Check OpenAI status when controlled files fail across clients or networks. At 16:29 Asia/Shanghai on July 12, 2026, the page reported fully operational systems. That timestamp does not disprove an earlier incident, a short-lived failure, or a problem limited to a tier, model, region, or feature; the status page itself describes availability as aggregate.
If there is an active upload incident, preserve your files and wait rather than consuming more attempts. When the status page is clear but the controls still fail across two official clients, prepare a compact support packet:
- account email, without sharing the password;
- timestamp and timezone for a controlled failure;
- browser or app name and version, operating system, device, and network type;
- a screenshot of the literal error;
- request ID, if the interface shows one;
- whether the known-good PDF and original each failed;
- whether the same test failed in a private session and on another network;
- a sanitized sample only if it contains no confidential data.
Contact OpenAI support through the Help Center chat. Do not send full payment-card data, passwords, API keys, or a confidential source document merely to prove that an upload failed. A reproducible small sample plus the control matrix usually gives support a cleaner starting point.
Frequently Asked Questions
Why does a PDF upload stop halfway and say “Unknown Error Occurred”?
The visible percentage only shows that part of the upload flow progressed; it does not identify whether later validation, parsing, quota, client, network, or service handling failed. Test a known-good PDF and the original in a fresh official client. Their outcomes are more diagnostic than the progress point.
Does ChatGPT support PDF uploads?
Yes. OpenAI's current supported-file guidance explicitly lists PDF. A generic unknown error is therefore not the same as “PDF is unsupported”; it requires a file, quota, client/network, or service diagnosis.
Is 10 MB the ChatGPT PDF upload limit?
No current official source checked on July 12, 2026 supports a universal 10 MB ChatGPT PDF limit. OpenAI's File Uploads FAQ stated a 512 MB hard per-file limit and a separate 2-million-token cap for text and document files on that date. Other rate, storage, Project, and context boundaries can still block a smaller file.
Should I flatten or print the PDF again?
Only after a known-good PDF works and the original fails in a fresh client. Printing or exporting a copy can rebuild structure or flatten troublesome interactive content, but it is a community-derived normalization test, not a guaranteed fix. Preserve and compare the original because forms, links, signatures, layers, or accessibility data may be lost.
Do scanned PDFs cause this error?
A scan is a plausible file-specific suspect, especially if the document lacks selectable text or has unusual encoding, but OpenAI does not publish scans as the official cause of this literal message. Prove the branch with the two controls. If the scan alone fails, try a fresh OCR-enabled export of a copy or upload only the needed pages.
Can failed attempts use my upload quota?
OpenAI says failed upload attempts can sometimes count toward the rolling upload-rate cap. That is why repeated blind retries can make diagnosis harder. Run each control once, record the result, and retry only after changing one variable.
When should I contact OpenAI support?
Escalate when controlled PDFs fail across two official clients, or when a live incident remains unresolved after service recovery. Include the timestamp, timezone, platform, screenshot, request ID if available, and the results of both controls. Keep passwords, full card data, API keys, and confidential documents out of the ticket.