ChatGPT File Uploading: Supported File Types, Maximum Size Limits, Upload Rules, And Document Reading Features

ChatGPT file uploads turn documents, spreadsheets, images, and structured data into working context for extraction, transformation, and analysis.

The practical experience depends on the file type, the enabled tools in the chat, and the plan-level features available to the workspace.

·····

Supported File Types Are Grouped By How ChatGPT Ingests And Interprets Them.

File support is best understood by separating formats that are read as text, formats that are executed as data, and formats that require visual interpretation.

Documents and text files are typically ingested through text extraction, which favors selectable text over scanned imagery.

Spreadsheets and delimited tables are typically ingested through a data-analysis workflow that treats cells and rows as structured inputs rather than prose.

Images are processed through vision capabilities, which can interpret visible content and, when appropriate, read embedded text.

........

Common File Categories And How They Are Typically Processed.

Category	Typical Formats	Primary Processing Mode	Typical Output
Documents	PDF, DOCX, DOC, RTF, TXT	Text extraction and retrieval	Summaries, quotes, sections, comparisons
Presentations	PPTX, PPT	Text extraction from slide content	Slide-by-slide notes, rewrites, restructuring
Spreadsheets	XLSX, XLS, CSV	Data analysis workflow	Calculations, pivots, charts, cleaned tables
Structured text	JSON, XML, HTML, MD	Text parsing and schema inference	Field extraction, validation, transformation
Code	PY, JS, JAVA, CPP, CSS, etc.	Text parsing with syntax awareness	Explanations, refactors, patches, reviews
Images	PNG, JPG, WEBP, GIF	Vision interpretation	Descriptions, transcription, visual Q&A

Some formats that are technically “supported” still behave differently depending on whether the content is machine-readable.

A PDF made of selectable text behaves like a document.

A scanned PDF behaves more like an image container unless the plan and toolchain provide visual retrieval for PDFs.

·····

Maximum Size Limits Combine File Weight, Token Budgets, And Type-Specific Caps.

Upload limits are not governed by a single number, because raw file size and readable content size do not scale the same way.

A compact PDF can contain millions of characters, while a large slide deck can be visually heavy but text-light.

Text-bearing documents are also constrained by token budgets, which cap how much text can be ingested from a single file even when the file is small in megabytes.

Spreadsheets and CSV files are constrained by size and shape, because wide rows and large cell blocks expand rapidly in memory during analysis.

........

Widely Applied Size And Content Caps For Uploads.

Limit Type	Common Cap	Applies To	What It Means In Practice
Maximum file size	512 MB	Most uploadable files	The upload is rejected if the file exceeds the hard size ceiling
Maximum tokens per text file	2,000,000 tokens	Text and document files	Only up to the token cap can be ingested from a single document
Maximum CSV size	~50 MB	CSV and similar delimited data	Large tables may require splitting or sampling for reliable analysis
Maximum image size	20 MB per image	Image uploads	Oversized images must be compressed or resized before upload

Token limits matter most for long PDFs, long DOCX reports, and large text exports.

When a document approaches the token ceiling, later sections may be partially ingested, skipped, or summarized less reliably than earlier sections.

When a dataset approaches the CSV size ceiling, splitting by date range, region, or table partition usually produces more stable results than forcing a single monolith.

·····

Upload Rules Depend On Where The File Is Attached And How The Workspace Is Configured.

Uploads can be attached inside a conversation, attached into a project workspace, or stored as knowledge inside a custom GPT, and each pathway carries its own caps.

Conversation uploads are optimized for short-lived analysis tied to the current thread, which keeps the model’s context aligned to the dialogue.

Project uploads are optimized for repeated reuse inside a bounded workspace, which keeps related files grouped and reduces repeated re-uploading.

Knowledge uploads for a custom GPT are optimized for retrieval and reference across sessions with that GPT, which makes them more persistent than ad hoc chat attachments.

........

Common Quantity Limits By Attachment Context.

Attachment Context	Typical Limit	Notes
Single conversation	10 files	Designed for focused, thread-specific work
Custom GPT knowledge	20 files per GPT	Intended for persistent reference by that GPT
Single upload action	10 files at the same time	Applies when adding many files in one action

Rate limits are also part of the upload rules, because the system constrains how many files can be ingested over a time window.

A typical ceiling is measured in files per few hours, with stricter limits for free tiers than for paid tiers.

When the system is under load, dynamic throttling may reduce the effective limit temporarily even if the nominal quota would otherwise allow the upload.

Storage caps also apply across time, because uploaded files count toward per-user and per-organization storage quotas.

When storage caps are reached, additional uploads fail until older files are deleted or retention policies expire them.

·····

Cloud-Linked Uploads Behave Like Normal Uploads But Add Permission And Link Constraints.

Cloud connectors allow importing files from services such as Google Drive, OneDrive, and SharePoint when enabled.

A cloud-sourced file still counts as an upload and is still bounded by the same file-size, token, and quota limits.

The main difference is that the connector must be able to fetch the file at the time it is attached, which requires stable permissions and accessible sharing settings.

Links that work inside one organization may fail across organizational boundaries if the sharing policy does not permit external access.

If a conversation is later resumed after a long pause, cloud-based files may need to be reattached if the system no longer retains the fetched copy.

·····

Document Reading Features Vary Between Text Retrieval, Data Analysis, And Visual Understanding.

Document reading is not one feature but a family of behaviors that depend on what the file contains and which tools are active.

For PDFs and DOCX files, the default reading behavior focuses on extracting and indexing text so that passages can be quoted, compared, and rewritten.

For spreadsheets and CSV files, reading is typically coupled to computation, enabling filtering, aggregation, joins, and basic statistical summaries.

For images, reading is visual, enabling interpretation of diagrams, screenshots, scans, and photographed pages, including text where legible.

A critical dividing line is whether images embedded in documents are treated as meaningful content or ignored.

In many setups, embedded images in documents other than PDFs are not interpreted as visuals, which means a slide screenshot inside a DOCX may be effectively invisible unless the image is uploaded separately.

For PDFs, some enterprise configurations add visual retrieval, which can treat images, charts, and scanned pages as first-class inputs rather than discarded decoration.

........

Typical Reading Behaviors You Can Expect By Content Type.

Content Type	What Is Reliably Read	What Often Needs Extra Handling	Best Practice For Accuracy
Text PDF	Headings, paragraphs, tables as text	Complex layouts, footnotes, multi-column flow	Provide page ranges and ask for quoted evidence
Scanned PDF	Visible text when visually interpreted	Low-resolution scans, skewed pages, faint ink	Upload higher-resolution scans or individual pages
DOCX	Body text and simple tables	Text inside images, complex text boxes	Export to PDF or upload key images separately
PPTX	Slide text and speaker notes when present	Text embedded as images	Provide the deck plus screenshots of critical slides
CSV/XLSX	Rows, columns, values, formulas as data	Merged cells, inconsistent headers	Normalize headers and split huge sheets
Images	Visible text and objects	Tiny fonts, compression artifacts	Crop to the relevant region and increase resolution

When a request depends on provenance, it is safer to ask for page numbers, section headings, and direct quotations rather than relying on paraphrase alone.

When a request depends on computation, it is safer to ask for the exact columns used, the filtering rules applied, and the definitions of metrics.

When a request depends on visuals, it is safer to provide clean crops of the exact figure, chart, or paragraph that matters.

·····

Output Quality Improves When The Prompt Mirrors The File’s Structure And Constraints.

Large documents respond better to scoped prompts that target specific chapters, headings, or page intervals.

Dense technical PDFs respond better when the task is broken into extraction first, then synthesis, then rewriting, because each step reduces ambiguity.

Spreadsheets respond better when goals are stated in terms of columns, keys, and expected outputs, because the model can validate intermediate results against the table shape.

Mixed-content files respond better when you explicitly identify what matters, such as “use only the table on page 7” or “ignore appendix C.”

When a file includes multiple languages, domain jargon, or unusual encoding, providing a short glossary in the prompt can reduce misinterpretation.

When a file includes confidential or regulated material, the safest operational posture is to minimize scope, avoid unnecessary uploads, and delete files when the work is done.

·····

Retention, Deletion, And Training Rules Influence How Long Uploaded Files Persist.

Uploaded files are generally retained alongside the conversation or workspace context they were attached to, subject to plan-specific retention behavior.

Deleting a conversation or removing a file typically triggers deletion within a bounded window, while certain limited exceptions can apply for security or legal reasons.

Files uploaded as knowledge for a custom GPT are designed to persist until that GPT is deleted or the knowledge is removed.

Enterprise and business offerings typically apply stricter isolation, and consumer offerings may allow content to be used to improve services depending on user controls and policy settings.

For operational certainty, retention-sensitive workflows should document the plan’s retention behavior and treat file deletion as a deliberate step in the process rather than an assumption.

·····

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

[datastudios.org]

·····