Do the Spreadsheet Work in Tabs You Are Not Using

The documented job is parallel work across tabs, in the background, without taking over the browser, on websites you allow. OpenAI describes that as a way to combine research, data entry, and verification across web apps in one task.

The official prompt is short enough to copy:

Compare the open product pages, collect the plan limits in a table, cite each source tab, and flag any differences that need a manual check.

The prompt does not say the table is a spreadsheet file, that every cell will be correct, or that the agent may leave the tabs you named.

The current extension guide names the product more broadly. You use the ChatGPT browser extension from the ChatGPT desktop app in Google Chrome, Microsoft Edge, Brave, Opera, or Vivaldi. ChatGPT can read or act on sites where you are already signed in. The guide’s examples are LinkedIn, Salesforce, Gmail, and internal tools. All five browsers support tab mentions and browser control from the desktop app. Chrome, Edge, Brave, and Vivaldi also support side chat. Opera does not. Opera tasks start in the desktop app.

Browser availability can depend on rollout and your workspace settings. Update the desktop app before you add another browser.

Hand the tabs over when four conditions are already true, because each one is in the docs.

The pages are open. The May prompt starts from open product pages. Mention an open tab when you want that page as context. Do not treat a closed bookmark as the same instruction.

The job is research, data entry, or verification across web apps. That is the use OpenAI names for the background behavior. A limit table with a citation and a flag for manual review matches that sentence. A redesign of the vendor’s checkout does not.

You can name the host. Prompts are based on the website host, such as example.com. A task that stays on two or three hosts you can list is easier to allow, and easier to decline, than a task that hops through ads and redirects.

You will review the sheet. The prompt itself asks for a manual-check flag. The extension guide says to treat page content as untrusted context and to review the website before allowing ChatGPT to continue.

Use the extension when the limits sit behind a login you already have. The guide says ChatGPT can use a plugin when a dedicated integration exists, your browser when the task needs signed-in browser context, and the built-in browser for localhost. If the same facts are already in a plugin or a file, use that path.

Decline the handoff when the page is the risk.

Page content, selected text, and video transcripts are untrusted context. Review the page and any requested permissions before you ask ChatGPT to use or act on that information. A product page that quotes a limit is still someone else’s page. A limit can be a promotion, a region, a cached banner, or a sentence about a different plan. The May prompt’s last clause exists for that reason: flag differences that need a manual check.

Do not send secrets or highly sensitive data through a browser task unless they are required and you are present to review each prompt. The extension can operate on sites where you are signed in. That includes mail and internal tools in the guide’s own examples. A sheet of public plan limits is a different risk from a sheet built out of an inbox.

Do not choose Allow for all sites to save a click. That option is marked elevated risk. ChatGPT no longer asks before using websites. The guide says to choose it only if you trust ChatGPT to use any website open in the browser. A comparison of three product tabs does not need every other tab you left open.

Do not use browser history as a stand-in for tabs you forgot to open. History access is elevated risk. It can include telemetry, internal URLs, search terms, and activity from signed-in devices. Relevant entries can enter the task, and misleading page content can raise the chance that data is copied somewhere you did not intend. ChatGPT asks each time. There is no always-allow for history.

Do not use the regular browser profile for a local app you are building. For that, the extension guide says to use @Browser. The built-in browser supports sign-in and keeps the browsing work inside ChatGPT, without your regular browser profile. The Computer Use guide says to use the built-in browser first for web apps you are building locally.

Chrome’s permission prompt may include access to the page debugger, reading and changing all your data on all websites, reading and changing browsing history on all signed-in devices, notifications, bookmarks, downloads, communication with cooperating native applications, and viewing and managing tab groups. Those permissions make browser workflows possible. They are not the same as a standing yes. ChatGPT still uses its own confirmations, settings, allowlists, and blocklists before it uses websites or browser history during a task.

By default, ChatGPT asks before it interacts with each new website. The prompt is based on the host. When it asks, you can choose:

• Allow once, so it can use the website one time.
• Allow for this site, so it can use the website again without asking.
• Allow for all sites, so it can use websites without asking.
• Decline, so it does not use the website.

Manage the lists in the desktop app under Settings, then Computer Use, then Manage next to the browser. The allowlist holds domains it can use without asking again. The blocklist holds domains it should not use. The supported browsers share these website permissions. Remove a domain from the allowlist and it asks again. Remove a domain from the blocklist and it can ask again instead of treating the domain as blocked.

Memories follow the Memories setting. If Memories is on, ChatGPT can use relevant saved memories while working in the browser. If Memories is off, browser control does not use memories. Turn memories off when the sheet should come from the open tabs and not from an earlier chat.

OpenAI does not keep a separate complete record of extension actions. It stores browser activity only when that activity is in the chat: page text, screenshots, tool calls, summaries, messages, or other included content. ChatGPT data controls apply to that context.

You stay in the foreground. The May note says the extension can do the tab work without taking over the browser. Keep the sheet, the prompt, and the tab you trust in front of you. Let the other product tabs sit behind.

Side chat does not follow you from tab to tab. The panel stays with the tab where you opened it. If you want the conversation beside the vendor page you are checking, open side chat on that tab. If you want the task in the desktop app, start a ChatGPT Work or Codex chat and @-mention the browser there. Both paths are documented. They are not the same window.

Keep these in your hands: the host list, the manual-check column, and the browser profile where the extension is installed. Allow the product hosts you named. Block ad hosts, login walls, and internal tools that happen to be open. A second profile without the extension is a setup miss, not a model miss. Keep localhost on @Browser. On macOS, a scoped Computer Use task can run in the background while you work elsewhere, after Screen Recording and Accessibility permissions. On Windows, Computer Use stays on the active desktop and takes the foreground. Do not plan that Windows app job as background tab work.

The sheet is a claim with citations, not a bill. Check it in the order the prompt already uses.

Open each cited tab. The instruction is to cite each source tab. A row without a tab is unfinished, even if the number looks familiar. If the agent cites a tab you did not have open, stop and look at the allowlist before you keep the row.

Read the limit on the page, not only in the cell. Plan pages mix seats, regions, annual billing, and promotional lines. The docs do not define a schema for “plan limit.” You do. Decide the columns before you send the prompt: plan name, billed period, what the page states, the tab, and whether you still need to check it. The agent’s table can be wider. Your check uses those five.

Compare rows that disagree. The prompt asks for differences that need a manual check. A disagreement between two tabs is the point of the task. A disagreement between the tab and the cell is a reason to drop the cell.

Highlighted text and, when captions exist, a YouTube transcript can enter the chat. Both are untrusted context. A transcript is not the pricing page.

The extension guide does not publish a per-tab price. A browser task starts in a ChatGPT Work or Codex chat. On the pricing page checked the same day, ChatGPT Work and Codex share usage. Work usage inside ChatGPT uses the same pricing, credits, and usage limits as Codex. API token prices are separate. Do not use them to estimate included tasks.

Plan prices on that page, in USD: Free is $0 per month, with GPT-6 Luna at Standard speed in the desktop app, subject to rollout. Go is $8 per month, with the same Luna note, subject to rollout. Plus is $20 per month and lists Codex on the web, in the CLI, in the IDE extension, and on iOS, GPT-6.1 Sol and GPT-6 Luna, and ChatGPT credits when you need more than the included allowance. Pro starts at $100 per month, with plans at $100, $200, and $500. Ultrafast mode is listed on Pro $500. Business is $20 per user per month billed annually for 2 or more users, or $25 per user per month billed monthly, and lists desktop and mobile access, larger cloud-chat machines, credits, and admin controls. By default, OpenAI does not train on Business data. Enterprise and Edu are contact-sales here, with no public seat price in the section checked. The card lists priority processing, admin and retention controls, and usage monitoring through the Compliance API.

For Plus and Standard Business, the pricing page prints local-message estimates per five-hour period. It says these are not fixed message limits. Pro plans currently have no five-hour limit. Cloud tasks may use more of the allowance than local messages. Local messages and cloud chats share the plan allowance. Weekly limits may also apply. Enterprise and Edu with flexible pricing have no fixed rate limits and scale with credits. Enterprise and Edu without flexible pricing have the same per-seat usage limits as Plus for most features.

The printed ranges, local messages per five hours, for Plus and Standard Business, are GPT-6 Astra 5–45, GPT-6.1 Sol 15–160, GPT-6.1 Terra 35–285, GPT-6 Sol 15–150, and GPT-6 Luna 350–3,000. Fast mode uses included usage at 2.5 times Standard, and purchased credits or Enterprise pay-as-you-go at 2 times. GPT-6 Astra Ultrafast and GPT-6.1 Sol Ultrafast use included usage at 8 times and those paid rates at 6 times. The extension pages do not assign a speed to a background tab task. Leave Standard unless you meant to pay the multiplier. If a turn hits the limit, that turn can continue, subject to fair use. Plus and Pro can buy credits. Flexible Business, Edu, and Enterprise plans can buy workspace credits. Credit prices depend on the plan. An API key can run extra local chats at API rates, in the CLI, SDK, or IDE extension, without cloud features such as GitHub code review. Check the usage dashboard for current limits. In the Codex CLI, /status shows what remains. The page suggests looking every week or two.

Illustrative only: if one tab pass were a single local GPT-6 Luna message at Standard speed, the printed low end is 350 such messages in five hours and the high end is 3,000. That division is not a quote. The ranges are estimates, cloud tasks can use more of the allowance, and similar tasks can consume different amounts. Read the dashboard after one real pass before you queue the next.

Update the ChatGPT desktop app first. Install the browser you intend to use. In the desktop app, open Settings, then Computer Use. If the browser is not in the main list, expand More browsers.

1. Select the browser and follow any prompt to install the required plugin.
2. Select Install beside the browser. That opens the extension store page. Install the ChatGPT extension and review the browser’s permission prompts.
3. Return to Computer Use and confirm the browser shows Manage.
4. Start a ChatGPT Work or Codex chat and select the browser with an @-mention. Use the profile where you installed the extension.

The guide’s example is @Edge open Salesforce and update the account from these call notes. For the clerk job, name the open product tabs, ask for plan limits in a table, require a source tab on every row, and require a flag where the pages disagree. Then set the hosts.

• Leave the default ask-before-each-new-host behavior in place.
• Allow once, or allow only the product hosts, for the tabs in the task.
• Put every other sensitive host on the blocklist if it is open in that profile.
• Leave Allow for all sites off.
• Leave history unused unless a later prompt asks and you still want it.
• Turn Memories off if the sheet should not pick up earlier chats.
• If the task must upload a local file in Chrome, turn on Allow access to file URLs, then start the task again.

Can the Chrome extension work in background tabs while I keep using the browser?

Yes. The What’s new entry for May 4–8, 2026 says the Chrome extension can work in parallel across tabs in the background without taking over your browser. You control which websites Codex can use. The current extension guide, checked 10 November 2026, still documents tab mentions, side chat on four of the five browsers, and host-by-host approval. It does not repeat the background sentence. This article relies on the What’s new page for that sentence.

What is the official spreadsheet example?

The published prompt is: compare the open product pages, collect the plan limits in a table, cite each source tab, and flag any differences that need a manual check. It is an instruction, not a sample file.

Which browsers can start this from the desktop app?

Google Chrome, Microsoft Edge, Brave, Opera, and Vivaldi. All five support tab mentions and browser control from the desktop app. Opera does not support side chat.

Does the extension ask before it opens a new site?

By default, yes. The prompt is per website host. You can allow once, allow that site, allow all sites, or decline. Allow for all sites is marked elevated risk because later sites are not asked. Allowlists and blocklists are managed under Settings, then Computer Use, then Manage, and the supported browsers share them.

Does a background tab pass have its own price?

No separate per-tab price is published. The task starts in a ChatGPT Work or Codex chat, and those surfaces share usage. Included-task estimates are not API token prices. Plus and Standard Business show local-message ranges per five hours, and the page says those ranges are not fixed limits. Read the usage dashboard after a real pass.

Should I use the same method on a local app or a Windows desktop app?

No. Local web apps you are building go to the built-in browser with @Browser. Windows Computer Use runs on the active desktop and cannot work in the background while you keep using that same Windows session.