Claude vs ChatGPT vs Gemini: Knowledge Base Customization

Vendor pages document different knowledge-base features and limits for Claude, ChatGPT, and Gemini, including paid Claude project RAG, while Chat Picker has no assistant test results.

Sources checked 2 Oct 2026

For Claude vs ChatGPT vs Gemini knowledge base customization, the vendor pages document project files and instructions for ChatGPT and Claude, while Google’s pages document uploads and Workspace-connected sources for Gemini; they do not provide a shared test of retrieval quality or refresh behavior. The features, limits, and controls are not documented as equivalent. Chat Picker has not tested the assistants for this use.

What the vendors document for Claude vs ChatGPT vs Gemini

Vendor figures below are as read on the linked pages in October 2026.

ChatGPT. OpenAI’s Projects page says Projects keep chats, files, and instructions together and can use that context for ongoing work; access depends on plan and workspace settings. OpenAI’s File Uploads FAQ covers common text, spreadsheet, presentation, and document extensions. It sets a 512MB hard limit per file, a 2M-token cap for each text or document file but not spreadsheets, an approximate 50MB CSV or spreadsheet limit, and 3 uploads per day for Free users.

Claude. Anthropic’s projects page says Projects have separate chat histories and knowledge bases for documents, text, and code, plus project instructions; Free users can create at most 5 projects. Enhanced knowledge with RAG is available only on Pro, Max, Team, and Enterprise, and can expand capacity by up to 10x near context limits. Anthropic’s upload page gives a 500MB chat limit and 30MB project limit. For PDFs up to 100 pages, Claude analyzes text and visuals; from 101 to 1,000 pages, it processes text only.

Gemini. Google’s file-upload page says Gemini Apps accepts documents, spreadsheets, notebooks, photos, videos, and more. A prompt can include up to 10 supported files, subject to availability; each video can be 2GB and every other supported file type 100MB. Paid AI plans can raise feature limits. Adding Drive files requires Keep Activity and a connected Workspace. The Google pages read here do not describe a project-level knowledge base or custom project instructions.

Plans and scale. ChatGPT Plus is $20/month. Claude Pro is $20 monthly or $17/month with annual billing and $200 upfront. Google AI Pro is $19.99/month. Heavy-use options include OpenAI Pro 100, Pro 200, and Pro 500 at $100, $200, and $500 per month; Claude Max from $100/month; and Google AI Ultra at $99.99/month or $199.99/month.

Data handling. OpenAI’s data-controls page says controls vary by sign-in, plan, and workspace; disabling “Improve the model for everyone” stops new chats from training OpenAI’s models but does not delete or hide saved chats. Anthropic’s model-training page says consumer chats and coding sessions may be used to improve Claude when the user allows it; connector content is excluded from that training data unless copied into a chat, while incognito chats are not used. Google’s Privacy Hub names Google Ireland Limited in the EEA and Switzerland and Google LLC elsewhere; the US plans page states no model-training setting.

Accuracy and policy. OpenAI’s accuracy note says ChatGPT can be incorrect or misleading, may sound confident when wrong, and does not include events beyond its knowledge cutoff unless tools are used; verify important information. Anthropic’s note says not to use Claude as a sole source and to review cited sources. Gemini source guidance says source links may cover websites, uploaded files, and connected Workspace documents or email. OpenAI’s usage policy bars unauthorized aggregation or distribution of private or sensitive information; Anthropic’s policy requires qualified professional review for covered advice, recommendations, and subjective decisions, plus AI disclosure for direct outputs; Google’s guidelines say Gemini should not generate factually inaccurate outputs that could significantly harm health, safety, or finances.

What the documentation cannot tell you

The pages read here do not give comparable details about vector databases, embeddings, chunking, reranking, or retrieval architecture. Claude’s page names RAG for paid projects, but the cited OpenAI and Google pages do not explain an equivalent architecture.

They also do not establish refresh latency after a file replacement, connector synchronization behavior, retrieval accuracy on a common corpus, or the compliance controls that apply to your deployment. Your trial can show behavior for your files and settings; it cannot establish a universal ranking. Chat Picker’s method publishes vendor-sourced comparison facts, not its own quality, speed, accuracy, reliability, or benchmark results.

How to check it yourself

Formats and parsing. Give each assistant the same synthetic PDF containing prose, a table, and an image, plus a DOCX, CSV, and spreadsheet. Ask: “Answer only from the attached files. In the Atlas operations handbook, on which weekday are returns reviewed? Give the file name, page number, and exact sentence. If absent, reply ‘Not found.’” Record accepted or rejected formats, missing visual or table content, citations, and limit warnings.
Retrieval. Ask: “Compare the returns rule and refunds rule in the attached Atlas documents. Cite the file and exact passage for each rule, then identify any conflict. If a rule is missing, say so.” Look for relevant passages, unsupported synthesis, and contradiction handling. Record omissions, incorrect citations, and whether the answer stays within the files.
Refresh. Replace the same-named handbook with a document saying returns are reviewed on Tuesdays. Ask: “Re-open the current Atlas source and answer: on which weekday are returns reviewed? Cite the current file and passage.” Look for stale text, duplicates, and refresh steps. Record the timestamp, steps required, and source version used.
Controls. Use a non-sensitive dummy file and ask: “List every source used for the Atlas answer. Label each as an uploaded file, connected Workspace content, or web content.” Inspect visible source links and account settings. Record the plan, account type, training controls, connector permissions, retention, deletion, export, and provider or region information shown to you.
Limits and scaling. Repeat the bundle and task under each candidate plan. Ask: “Using only the attached Atlas files, answer the returns question and state which files or sections you could not access.” Record upload or usage limits, context warnings, plan notices, price, billing basis, and the vendor page’s read date.

Which rows of the comparison matter

Use the Claude vs ChatGPT vs Gemini matrix as a checklist. Start with “Free plan,” “Low-cost tier,” “Main paid plan,” “Heavy-use plans,” “Team plan,” “Context window in the app,” “How usage limits are described,” and “Training on your chats.” For project knowledge and ingestion, return to the linked vendor pages above.

Treat “Not verified” and blank cells as unknowns, not zero. Recheck the vendor page before subscribing because prices differ by country and plans change.