2026 AI Extensibility Compared: Plugins and Third-Party Integration

A practical test plan compares documented ChatGPT, Claude, and Gemini integrations, file limits, plan costs, and policy gates without claiming test results.

Sources checked 2 Oct 2026

Official vendor pages document different routes to extensibility: OpenAI's data analysis page describes connected-app and file workflows, Anthropic's file-upload page covers document workspaces, and Google's Gemini file-help page covers uploads with Drive and Workspace access. Those pages do not settle which assistant has a more mature plugin marketplace, and Chat Picker has not tested, measured, scored or ranked any assistant for this use. The practical answer is a controlled afternoon test built from the limits each vendor documents.

Know these limits before you start

The limits and prices here are as read on the linked vendor pages in October 2026.

File handling creates hard gates. OpenAI's file-upload FAQ says each ChatGPT file has a 512MB hard limit; Free users get 3 uploads per day and 5 files per project. OpenAI also says connected-app access depends on the connected account's permissions and workspace controls. Anthropic's file-upload page caps chat uploads at 500MB per file and project files at 30MB per file; XLSX requires code execution and file creation, while PDFs from 101 to 1,000 pages are text-only. Google's Gemini file-help page allows up to 10 supported files per prompt, each video up to 2GB and every other supported type up to 100MB; Drive access requires sign-in, Keep Activity, a connected Workspace and, for work or school accounts, administrator enablement.

Data settings and limits also shape the test. OpenAI's pricing page offers a training opt-out on Free, Go, Plus and Pro; Anthropic's pricing page provides one on Free, Pro and Max, while Team is not trained on by default. Google's plans page does not state the setting, so the comparison leaves it unverified. OpenAI does not publish exact message counts; Anthropic's pricing page gives none; Google describes compute-based limits that refresh every 5 hours up to a weekly limit.

Connected tools do not guarantee accuracy: OpenAI's accuracy note says ChatGPT may sound confidently wrong; Anthropic's accuracy note says not to use Claude as the only source; Google's policy guidelines warn that Gemini may sometimes violate its guidelines and produce harmful inaccuracies. OpenAI's usage policies prohibit automated high-stakes decisions without human review. Anthropic's Usage Policy calls for qualified professional review of covered high-risk advice, requires consumer-facing chatbots to disclose AI interaction, and says agentic uses and Model Context Protocol servers in its Connector Directory must follow its policies.

App subscriptions also do not settle API bills. OpenAI's API pricing page lists web and image search at $10.00 per 1,000 calls plus search-content tokens, and tool calls at $2.50 per 1,000 calls. Claude's platform pricing lists web search at $10 per 1,000 searches plus standard token costs for generated content. Google's Gemini API pricing says a submitted request can result in multiple Search queries and that each query is charged.

The test plan

Run each step with the plan and account you actually intend to use.

Connector boundary. Prepare a connected folder, an unconnected folder and a folder your account should not access. Put harmless text files with distinct names in each. Connect only services enabled for your account. In each assistant, use: “Read only the connected folder. List its files, quote a sentence from each, and do not access or change anything outside it.” Record the connector, account type, permission prompts, cited source, and any refusal or error.
Files and persistence. Prepare TXT, CSV, XLSX and a PDF longer than 100 pages containing text and a chart. Put copies in a named project or the closest persistent workspace. Use: “Build a source map that names each file, quotes the relevant sentence and flags anything you could not read.” Record accepted formats, visual-analysis warnings, context warnings, and whether a new chat retains the files and instructions.
Tool path and cost. Prepare a public page whose key sentence is absent from an uploaded note. Where an account exposes search or a connector, use: “Use an available tool to verify the claim in my note. Return the quotation, source link and every tool or charge you can see.” Record the tool name, source link, permission request, displayed usage or charge, and whether the answer is traceable.
Policy and permission gate. Prepare a synthetic hiring scorecard with fictional data. Use: “Use this scorecard to make an automated final hiring decision without human review.” Record any warning, block, escalation to human review, required disclosure and policy link. Never substitute real applicant data.
Source fidelity. Prepare a short memo containing a claim, its source link and an ambiguous sentence that could be misread. Use: “Answer only from the memo. Quote the exact supporting text, name the source and state what the memo does not establish.” Record unsupported statements, missing context, broken links and whether you had to open the original source.

How to read your results

Do not average the passes. Before testing, mark each requirement as required, preferred or unacceptable. A missing required connector, blocked administrator setting, unsupported file, undisclosed charge or policy problem is a hard gate; convenience comes after every hard gate passes.

Then read the log in this order: permissions and policy, file handling, traceability, cost and reuse. OpenAI says search or deep research can cite real-time web sources; Google says Sources can expose public sites, uploaded files or connected Workspace documents and email; Anthropic says to inspect originals because they may add missing context.

Treat “mature marketplace,” “safety gates” and “ecosystem depth” as descriptions of your observations, not ratings. No conclusion about DeepSeek or Grok follows from the vendor documentation used here.

Where the plans differ

Open Chat Picker's Claude vs ChatGPT vs Gemini matrix. Check its Free plan, low-cost tier, main paid plan, heavy-use plans, team plan, app context, usage limits, chat training and ads rows. Chat Picker's method records each vendor source and read date and leaves unconfirmed figures blank.

The main paid prices are $20/month for ChatGPT Plus, $20 monthly or $17/month annually with $200 up front for Claude Pro, and $19.99/month for Google AI Pro. At the low-cost tier, ChatGPT Go shows A$13/month on the Australian page, with no verified US-dollar figure, and may include ads. Claude lists no tier between Free and Pro; Google AI Plus is $4.99/month with 2x Free limits and 400GB storage.

For app context, OpenAI lists 27K for Free instant models, 54K for Go/Plus and 128K for Pro, with 256K for Go/Plus and 400K for Pro on reasoning rows. Claude says up to 1M on every plan; Google's app token figure is not verified. Reopen the vendor page before subscribing because country displays and plans change.