ChatGPT takes its place in Chrome and can now act on open tabs

ChatGPT expands its Chrome integration: multi-tab analysis, forms, downloads, and website actions from Work.

Comparing multiple offers open in Chrome, filling out a form using a document, or retrieving information spread across different pages. ChatGPT for Chrome is evolving with deeper integration into Google's browser. The extension now allows users to check multiple tabs to answer a question and, through Work mode, perform certain actions directly on websites. OpenAI also announced an enhanced interface providing access to chat history without leaving the current page. Open tabs become a source of context

The first evolution concerns ChatGPT's ability to leverage pages already open in Chrome.

Previously, using an assistant to compare multiple sites often required submitting their URLs, copying content, or repeating a search.

With the new integration, ChatGPT can examine the tabs to which the user grants access and cross-reference their information.

For example, someone browsing several online stores can request a comparison of dimensions, materials, delivery times, and return policies for different products.

The same principle applies to software offerings: ChatGPT can browse multiple pricing pages and verify which plans match a list of requirements.

The analysis can also cover articles, documents, or service providers presented on different sites.

This feature is accessible from Chat mode, which is dedicated to conversations and information analysis. It does not require granting the assistant permission to execute actions on pages. Work mode can use websites

The second evolution concerns ChatGPT Work, which allows users to delegate tasks involving multiple operations.

When used with Chrome, Work can interact with open pages rather than being limited to reading their content.

The agent can notably fill in fields, click on elements, organize tabs, download files, or draft messages in services accessible from the browser.

OpenAI gives the example of an event registration form. The user can ask ChatGPT to retrieve information from a document, complete the corresponding fields, and pause before submitting.

The same process can be used to review multiple suppliers, gather their specifications, and enter the results into a comparative spreadsheet.

These operations are performed within the context of the visited sites and the accounts to which the user is logged in, subject to the necessary permissions.

The difference between the two modes is therefore functional: Chat leverages tabs to inform its responses, while Work can act on interfaces. Comparing information then preparing an action

The integration combines information retrieval with operations performed based on those results.

OpenAI specifically demonstrates a scenario where several home renovation companies are open in different tabs.

ChatGPT can compare their services, identify the one that best matches the provided criteria, and draft a contact message.

Another example involves sales data. The agent can review multiple vendor pages, extract pricing, key features, and source links, then report them in a comparison table.

For a registration or administrative process, Work can also use information already present in a document to pre-fill a form.

The user can instruct the agent to pause before any submission, in order to verify the information and retain the final decision.

These examples show how multiple pages can be used within a single task, from collecting information to utilizing it in another application. A ChatGPT interface directly in Chrome

OpenAI is also updating its extension's interface.

ChatGPT can now be used alongside a website, without forcing the user to constantly switch between the browser and a separate chat window.

The interface provides access to the full ChatGPT conversation history, allowing users to find an existing exchange or continue ongoing work while browsing.

Both Chat and Work modes are accessible from this environment.

A user can thus read an article and ask for an explanation, then switch to a task requiring interaction with the site or other tabs.

The goal of this setup is to keep the visited page as the primary workspace, while ChatGPT assists alongside to search, analyze, or perform operations. Organizing tabs and downloading files

Tab control is not limited to page content.

In Work mode, ChatGPT can also assist with their organization, notably by grouping open tabs around the same topic.

This feature can be used to reorganize a research session containing numerous pages, without having to perform each grouping manually.

The agent can also initiate downloads from sites it has access to.

These capabilities are useful for tasks requiring the collection of documents, references, or multiple files from different services.

The actual capabilities remain dependent on site interfaces, access restrictions, and the permissions granted to ChatGPT. Permissions to access sites and perform actions

Using Chrome tabs introduces privacy and control considerations.

The extension can work with open pages and, in some cases, active authenticated sessions. Granted permissions therefore determine the information the assistant can access and the operations it can perform.

OpenAI includes access controls for websites and permission prompts depending on the context.

Users can notably review the affected pages and require human verification before a binding action, such as submitting a form.

Content from visited sites must also be treated as external sources that could contain misleading or malicious instructions. Action monitoring remains particularly important when the agent uses professional accounts or handles confidential information.

Site access also does not mean that all of its features are automatically compatible with ChatGPT's actions. An extension announced for all subscriptions

The new version of ChatGPT for Chrome is available on the Chrome Web Store.

OpenAI indicates that the extension is accessible to users across all ChatGPT subscription plans.

However, this general availability does not mean that all features share the same limits or permissions. The