AI Assistant File Processing: Formats and Compatibility Checks
Compare vendor-documented formats, limits, and plan constraints, then test ChatGPT, Claude, and Gemini on the same files in an afternoon.
Sources checked 2 Oct 2026
AI assistant file processing is not one compatibility switch: OpenAI’s File Uploads FAQ, Anthropic’s upload-files page, and Google’s Gemini file-upload page document different formats, caps, plan rules, and PDF handling. Those pages state vendor limits, but Chat Picker has not tested ChatGPT, Claude, or Gemini on your files.
Know these limits before you start
As read on October 1, 2026, OpenAI’s File Uploads FAQ sets a 512MB hard limit per file, a 2M-token cap for text and documents but not spreadsheets, approximately 50MB for CSV and spreadsheet files depending on row size, and 20MB per image. Its stated Free allowance is three uploads per day, subject to peak-hour reductions and other plan or account settings.
Read on the same date, Anthropic’s Upload files to Claude page sets 500MB per chat file and 30MB per project file, with possible extra token limits. XLSX requires code execution and file creation. Claude lists PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON, XLSX, JPEG, PNG, GIF, and WebP. It analyzes text and visual elements in PDFs of 100 pages or fewer, but only text from 101 to 1000 pages, and cannot read embedded images in non-PDF documents.
Read on that date, Google’s Upload & analyze files page says sign-in is required and Gemini supports most file types, with up to 10 files per prompt subject to availability; every file inside a folder counts. A video can be 2GB and each other supported file 100MB. Total video length is five minutes, or one hour with Google AI Pro or Ultra; total audio length is 10 minutes, or three hours on those plans.
Data settings belong in the log. Read on that date, OpenAI’s upload FAQ says chat files remain in the account until deletion and associated files are removed within 30 days after the related chat, account, or custom GPT is deleted, subject to exceptions; Google’s file-upload page requires Keep Activity and a connected Workspace for Drive files.
The ChatGPT, Claude, and Gemini pricing pages, also read on October 1, 2026, list training opt-outs on ChatGPT Free, Go, Plus, and Pro and Claude Free, Pro, and Max; Claude Team is not trained on by default. The Gemini plans page states no training setting, so Chat Picker marks it “Not verified.”
Use synthetic or properly redacted files. OpenAI’s usage policy bars unauthorized aggregation of private or sensitive information; Anthropic’s Usage Policy classifies legal guidance as high-risk and requires qualified professional review before finalization or dissemination; Gemini’s safety guidelines warn against factual inaccuracies that could cause significant harm.
Accuracy checks still apply. OpenAI says ChatGPT can sound confident when wrong, Anthropic says not to rely on Claude as the sole source of truth, and Google says displayed sources may include uploaded files and connected documents. Inspect the original rather than treating a source panel as proof.
The test plan
Create one answer key, use a fresh chat per assistant, and keep files, prompts, and recording rules identical across products.
-
Structured and scanned PDFs. Prepare a born-digital PDF, a scan-only PDF, and a mixed PDF containing a table and chart. Give each assistant: “Return a page-by-page extraction table. Use the page numbers shown in my PDF viewer, quote each field, identify unreadable text, and cite the page for every value. Do not infer missing text.” Record acceptance, missing fields, table errors, chart values, citation problems, and elapsed time.
-
Spreadsheets. Prepare a multi-sheet XLSX and a matching CSV with formulas, mixed data types, blank cells, and totals. Ask: “Inspect both files. List every sheet, preserve headers and data types, calculate the totals, and flag every value or formula that differs between the workbook and CSV. Show the row location for each discrepancy.” Record missing sheets, type changes, formula differences, total mismatches, rejection messages, and elapsed time.
-
Images. Prepare a labeled chart, an interface screenshot, and a handwritten note. Ask: “Read all three images. Return exact visible labels and numbers, transcribe the handwriting, and mark any character you cannot read. Separate visible evidence from interpretation.” Record omitted labels or values, uncertain handwriting, visual dependencies, errors, and elapsed time.
-
Audio and video. Prepare clean and noisy audio plus a video with on-screen text and an audible change. Ask: “Transcribe the audio and video. Add timestamps for speaker changes and on-screen text, identify audible or visible events, and mark unclear words instead of guessing.” Record rejection, truncation, speaker labels, timestamp alignment, missed text, and elapsed time.
-
Corrupted and unusual files. Prepare a password-protected PDF, encrypted ZIP, empty file, misleading file extension, and unsupported format. Ask: “Inspect this file only at the file-format level. State whether you can read it, name any blocker such as encryption, corruption, or an unsupported format, and do not bypass protection or invent contents.” Record the exact message, whether the file was rejected, whether a clean export changes the outcome, and any workaround required.
-
Compatibility and speed. Repeat the same file-and-prompt task in a fresh chat on the plan you intend to evaluate, using the same device and network where possible. Ask: “Name each file used, state any limit encountered, and separate direct observations from inference.” Start the timer when you submit and stop when the requested output is complete. Record the plan, upload route, file type and size, retries, partial output, and elapsed time. Do not combine unlike files into one speed result.
How to read your results
Turn the log into three checks: eligibility, fidelity, and operations. Eligibility means the file opens under your actual plan and data rules. Fidelity means required fields match the answer key, omissions are marked, and citations point to the right locations. Operations records retries, workarounds, and elapsed time.
Use pass, hold, or fail for each requirement, not an overall quality score. Remove a plan that misses a required format or data condition; among eligible plans, choose the one that meets your must-haves with acceptable logged effort. A documented protected-file rejection is an expected compatibility outcome, but unexplained failure on a known-readable file is a compatibility problem. This is your test, not a Chat Picker score.
Where the plans differ
Use the ChatGPT vs. Claude vs. Gemini comparison matrix and read the rows labeled Free plan, Low-cost tier, Main paid plan, Heavy-use plans, Team plan, Context window in app, How usage limits are described, Training on your chats, and Ads. Compare the file caps and PDF rules above against those plan entries, then check the linked vendor page before subscribing.
Pricing figures read on October 1, 2026, are $20 per month for ChatGPT Plus, billed monthly; $20 monthly or $17 per month on annual billing with $200 up front for Claude Pro; and $19.99 per month for Google AI Pro. Chat Picker’s method leaves an unconfirmed figure blank and marks it “Not verified” rather than filling the gap.