OpenAI
Updates

Current-state record · File workflows

What can ChatGPT do with files?

Short answer: OpenAI documents file-creation and review workflows. Below, we separate those descriptions from one observed Codex-assisted local run with downloadable spreadsheets, slides and PDFs. ChatGPT app behavior and universal reliability remain untested.

Inspect the downloadable evidence ↓

01 Evidence

A claim is only as strong as its label.

●

Officially documented

OpenAI describes the workflow. This is not an account-access check.

◆

Observed local workflow

Codex authored and checked three examples using Python libraries. Files, method and limits are available below.

○

Still unresolved

No ChatGPT UI test, cross-plan study, independent human review or model success-rate estimate.

02 Surface by surface

The interface changes the answer.

Documentation claims—not performance results. Availability can vary by account and rollout.

ChatGPT desktop app

Previews generated documents, presentations, spreadsheets, and PDFs. HTML previews may be available; supported previews can be annotated.

Official

ChatGPT Work on the web

Accepts source files and can create documents, presentations, spreadsheets, or PDFs for review, download, and revision.

Official

Codex CLI

Creates and edits working-directory files without a visual preview or annotation interface.

Official

Codex IDE extension

Creates and edits workspace files; compatible viewers handle documents, presentations, spreadsheets, and PDFs.

Official

Observed workflow / downloadable evidence

Three files.
One inspectable run.

Codex authored and ran a local Python workflow with invented data, then reopened the files and checked one revision per format. This is an observed tool-assisted example—not independent human review, a ChatGPT app test, or a model benchmark.

16 automated checks passedRun: Python 3.11.9 · Windows

01 / Spreadsheet

Do the formulas survive a revision?

Three invented order lines total 115. Changing the folder quantity from 3 to 5 gives an expected total of 155. Cell formulas and unaffected inputs were checked after reopening.

Observed limitation: formula caches are empty. The script checks arithmetic separately; it does not run Excel’s calculation engine.

Original XLSX ↓Revised XLSX ↓

02 / Presentation

Can one slide change without rewriting the others?

A three-slide brief changes “5 users” to “8 users” on the last slide. Reopened text confirms the requested replacement and unchanged first two slides.

Not tested: PowerPoint rendering, font substitution, speaker delivery or complex charts. Text preservation is not visual fidelity.

Original PPTX ↓Revised PPTX ↓

03 / PDF

Does a revision preserve the first page?

A two-page order summary replaces a status on page two. Extracted text confirms the total and replacement while page one remains unchanged.

Boundary: simple text and page checks, plus rendered previews. No OCR, tagged-PDF accessibility audit or arbitrary-file reliability claim.

Original PDF ↓Revised PDF ↓
Method, checks & reproduction files

One authored example and one targeted revision per format; automated structural/content checks. The same Codex session authored and checked the workflow; no independent human review.

Libraries: openpyxl 3.1.5 · python-pptx 1.0.2 · PyMuPDF 1.28.2. Model identity was not independently recorded; no model comparison is claimed.

  1. Workbook sheets survive save and reopen
  2. Spreadsheet revision retains line formula
  3. Spreadsheet independent arithmetic after revision
  4. Negative control: incorrect spreadsheet formula rejected
  5. Limit observed: no evaluated formula cache
  6. Presentation slide count
  7. Presentation titles survive reopen
  8. Presentation revision leaves slides 1 and 2 unchanged
  9. Presentation requested replacement
  10. Presentation unaffected title
  11. PDF page count
  12. PDF revised page count
  13. PDF total is extractable
  14. PDF revision leaves page 1 unchanged
  15. PDF replacement is extractable
  16. PDF prior status absent from revised page

Inspect the code before running it. Install the pinned requirements in an isolated environment, keep the script and fixture together, then run python run.py --output ./my-file-results. A reproduction proves the script’s behavior, not a fresh model attempt.

Python script ↓Synthetic input ↓Requirements ↓Full report and output hashes ↓
Rendered revised PDF page one, showing three synthetic order lines and total 115.00
Actual revised PDF / page 1
Rendered revised PDF page two with Revised example status and synthetic fixture identifier
Actual revised PDF / page 2

Evidence created by OpenAI Updates Editorial Desk with Codex assistance. Automated checks by the same workflow; no independent human sign-off. The files contain synthetic data only.

03 Not settled

Answers we will not pretend to have.

01

One limit for every plan

The source does not establish universal upload, page-count, or usage limits.

02

Equal availability

Rollouts and workspace settings can change what an account sees.

03

Consistent output fidelity

One scripted example does not establish correct layout, formulas or edits for arbitrary inputs.

04

A ChatGPT app test

No ChatGPT web or desktop session was exercised. The local examples do not validate previews, annotations, uploads or plan limits.

04 Source record

One baseline. Clearly bounded.

Official OpenAI documentation

Work with files

Reviewed 11 September 2026. It supports the surface claims above, not universal limit or reliability claims.

Read the official source ↗

05 Change record

Evidence, added in place.

— Documentation baseline established, then expanded with one Codex-assisted local workflow for three file formats. The report records the actual run time. ChatGPT app testing remains pending. Choose a working surface →