Browser and computer-use agents next 6 months: 10 products, one date (Nov 2026–May 2027)

OpenAI’s Computer Use guide says the ChatGPT desktop app can see and operate graphical apps on macOS or Windows, with ChatGPT Work and Codex. On Windows it takes the foreground of the active desktop. On a Mac it can keep working after the screen locks, and only if you turn locked use on. The guide does not mention Linux. It also does not add a computer-use fee on top of the plan. Plus is $20 a month. Pro is $100, $200, or $500 a month.

Cloud browser is a different machine. It is ChatGPT Work’s own browser, on paid plans except Free and Go. It can keep going after you close your computer, and it pauses for a sign-in or for an action that is hard to undo, including a payment. The password you type in the secure form is not shown to the model. It does not use the cookies already stored in your own browser.

A dot is the third OpenAI surface. The help page says it has its own cloud computer, that local access starts turned off, and that a connected computer can use your local browser when the cloud browser is blocked. Enterprise, Edu, and Healthcare workspaces are in an admin preview from November 4 through November 18, 2026. On November 18, billing starts and dots become available by default unless an admin opts out. The page describes flat rates or scaled consumption and does not print the dollars. Pro users have to be 18 or older, and the rollout excludes the European Economic Area, Switzerland, and the United Kingdom. Dots are not available for workspaces with UAE data or inference residency.

Claude’s help center says desktop computer use is in beta for Pro, Max, and Ultra, in Cowork and Claude Code, on macOS and Windows. Connectors come first, then a browser, then the screen. Team and Enterprise do not have it. The computer has to be awake and the desktop app has to be open. On macOS 15 or later, Claude can work in background windows. The API toolset is separate. computer_toolset_20260801 needs no beta header, gives Claude 17 member tools, and is not available in Claude Managed Agents. Declaring it adds about 4,500 input tokens to a request, about 4,590 on Claude Sonnet 5 and about 4,520 on the other supported models. There is no separate computer-use fee. You pay the tokens, and you run the screen.

Google’s Cloud computer-use page, last updated November 7, 2026, lists browser, mobile, and desktop. The default is the browser. Your code receives the click and runs it, usually with Playwright. The named models include Gemini 4 Argon and the Gemini 3.5 through 3.8 Flash line, plus Gemini 3 Flash preview and Gemini Robotics ER 2 preview. The page does not add a computer-use price on top of the model. Chrome auto browse is the consumer product. Google introduced it on January 28, 2026 as a U.S. preview for AI Pro and Ultra. The plans page read November 9 says it is available to Pro and Ultra users in the US and India. Google One’s restrictions add age 18 or older, Chrome 144 or newer, and a signed-in browser. Pro is $19.99 a month. Ultra 5x is $99.99 a month. Ultra 20x is $199.99 a month. The plans table leaves Google AI Plus blank for auto browse.

Comet is Perplexity’s browser. The product page lists Mac, Windows, iOS, and Android and does not print a price. The pricing page displayed US$0 a month and US$17 a month. The text that came back did not put the word Pro next to $17, so this report does not label that figure. Max, last modified September 14, 2026, is $200 a month or $2,000 a year, and it says Max has the highest weekly limit on Comet browser-agent queries. The weekly number is not on the page.

Browserbase sells the browsers, not a chat product. Free is $0, with 1 browser hour, 3 concurrent browsers, and 3 agent runs. Developer is $20 a month, with 100 hours and then $0.12 an hour. Startup is $99 a month, with 500 hours and then $0.10 an hour. Scale is custom. Stagehand, the SDK with act, extract, and observe, is Browserbase’s. Browser Use is the other infrastructure price that was on a public page: $0.02 per browser hour, billed by the minute with a one-minute minimum, no subscription, and a home-IP proxy at $5 a gigabyte. The library is MIT. The pricing FAQ says $100 in total payments unlocks 50 browsers at once, $1,000 unlocks 250, $5,000 unlocks 500, and $25,000 unlocks 1,000. A one-time $1 signup credit is mentioned. The hero on the same page pairs dollar amounts with concurrency in a layout this read could not resolve, so the free-tier concurrency is left unknown.

OSWorld 2.0 is 108 long computer-use tasks across 31 self-hosted websites. The page says a skilled person takes a median of about 1.6 hours, and that 69.6 percent of the tasks take more than an hour. It also says Claude Opus 4.7, at maximum thinking, averages 318 tool calls, against about 30 in OSWorld 1.0. Under the primary binary metric at 500 steps, Claude Opus 4.8 with maximum thinking and batched tool calls completes 20.6 percent of tasks, at a 54.8 percent partial score. The findings section puts that run at 244,000 output tokens, Opus 4.7 at 18.2 percent and 150,000 tokens, and GPT-5.5 at about 14 percent and 39,000 tokens. Above 163 minutes, binary completion falls to zero for every model kept in the plot.

Those sentences are what this table copies. An interactive chart on the same page did not return rows in the read on November 9, so a higher score that lives only in that chart is not copied here. The older OSWorld project page says OSWorld 2.0 has been the current version since June 26, 2026. Its verified score table did not come through as text, so this report does not reprint a Verified leaderboard.

Browser Use’s WebVoyager post is dated December 15, 2024. It reports 89.1 percent on 586 tasks with gpt-4o. The authors say they removed 55 tasks, changed years in tasks that had gone stale, and manually corrected evaluations the automatic judge marked unknown or failed. That is a real post. It is not a current public score for the product in the table above, and it is not the original WebVoyager task count. No vendor page read on November 9 printed a current WebArena score, so WebArena is not given a number here.

November 18, 2026, and the products already shipping. Nothing else has a day.

The dot help page is the only forward date. Enterprise dots are free to preview from November 4 through November 18, 2026. Starting November 18, billing begins for every enabled dot, and dots become available by default unless an admin opts out. Work a dot hands to ChatGPT Work or Codex is billed at the ordinary rates. Each member can create one personal dot. Texting is in beta for a limited number of Pro users in the United States, and it is not in Business or Enterprise workspaces. The page says a dot cannot call you at launch, and that it cannot be given its own email address at launch.

Everything else in the table is already in market or already in a named beta. That includes desktop Computer Use, the cloud browser, Claude’s desktop beta, the Claude API toolset, the Gemini computer-use API, Chrome auto browse, Comet, Browserbase, and Browser Use. “Already shipping” is not a promise about a later country, a later plan, or a later price. Those sit under Watch for when the page refused a date.

The undated rollouts can still land inside this window, because the pages have not closed them. That is a reading of their wording. It is not a forecast that they will.

Pro dots and Business Premium dots are “rolling out gradually,” and “access may take several days.” Read on November 9, that sentence still covers the rest of this window if a given account does not have a dot yet. It does not prefer December over May.

Texting “isn’t available for your account yet” is the help page’s instruction when the control is missing, with “try again later.” Later is not a month. U.S. Pro is the only group named.

Chrome’s January 28, 2026 post said Personal Intelligence would come to Chrome “in the coming months.” That phrase is more than nine months old. The plans page read on November 9 lists early access to Personal Intelligence in AI Mode in the United States, which is Search, not a sentence about Chrome. This report does not treat the January phrase as a Chrome ship date inside the window. It is recorded so a later check can see whether Google replaced it.