ChatGPT Dots vs Gemini Spark vs Meta Muse: AI Agent Browser Tests, Costs and Privacy Risks

That is the pitch, anyway. To see whether the latest crop of consumer agents has finally outgrown the tendency to stall out on basic web forms and hallucinate tasks, PCMag software writer Ruben Circelli recently put three major contenders—OpenAI's ChatGPT Dots, Google's Gemini Spark, and Meta's Muse—through an identical five-part gauntlet of everyday digital chores.

The results offer a sobering, practical reality check for Windows and desktop web users considering handing over their browser sessions, email inboxes, and credit cards to an automated assistant. While ChatGPT's Dots emerged as the only contender that consistently finished the job, the testing—paired with Google's own technical documentation on how desktop browser automation actually behaves under the hood—shows that clunky execution, authentication roadblocks, and serious privacy trade-offs remain the norm.

Meet the Contenders: Architecture, Models, and Price Tags​

Unlike traditional chatbots that stay fenced inside a text box, AI agents are built to take action across external services. However, according to PCMag's testing and official platform documentation, these three tools take noticeably different architectural paths—and charge wildly different admission fees—to get there.

The architectural divide here matters enormously if you are running these tools on a desktop PC. Both ChatGPT's Dots and Meta's Muse execute their web tasks inside isolated virtual browsers hosted in the cloud, alongside direct app connectors.

Google's Gemini Spark—which Google officially introduced in May 2026 running on Gemini 3.5 and its "Antigravity" harness—is designed to work both in the cloud and directly inside your local desktop Google Chrome installation. While Google has rolled out dedicated Model Context Protocol (MCP) connections for third-party services like Canva, OpenTable, Instacart, Dropbox, and Zillow since mid-2026, its standout desktop trick is "Chrome auto browse." Instead of spinning up a fresh virtual machine in a distant data center for every web errand, Spark can pilot the actual Chrome window sitting on your desktop.

Section Summary: OpenAI, Google, and Meta all promise hands-off task automation, but they split sharply on price and execution. Dots demands a steep $100-per-month minimum for consumer Pro access and works in a virtual browser, Muse offers a free entry point in a cloud browser, and Spark starts at $20 per month while attempting to drive your local Chrome session directly.

The Five-Task Gauntlet: Where Agents Succeed and Where They Stall​

To test whether these tools could handle real life, Circelli gave all three agents the exact same five assignments. Here is how the digital interns fared when asked to do actual work.

1. Grocery Shopping on Instacart​

The prompt was specific: locate J. Kenji López-Alt’s no-stir risotto recipe, research the best versions of each required ingredient stocked at a local Wegmans supermarket, and load them into an Instacart shopping cart.

• ChatGPT Dots: Passed smoothly. Once the tester linked his Instacart account to ChatGPT, the Dot located the recipe, matched the ingredients, and populated the cart without a hitch.
• Meta Muse: Finished the race, but limped across the line. Because Muse lacked a direct Instacart integration, it fell back to its virtual browser. That forced the user to manually type in an Instacart username and password, retrieve a two-factor verification code from email, and feed it into the session. Worse, Muse decided to toss expensive saffron threads into the cart—an ingredient the recipe never asked for.
• Gemini Spark: Failed outright. Even though Google began rolling out native Instacart connected-app support for Spark in July 2026, PCMag reported that connecting the app through Spark's settings page failed to work during the test. Spark found the recipe and suggested ingredients, couldn't access the account, and then froze while attempting to take control of the local Chrome browser.

2. Hunting for a Job via Gmail​

For the second test, all three agents were connected to the reviewer's Gmail account and told to locate his resume inside the inbox, then hunt down relevant job openings.

• ChatGPT Dots: Found the resume and returned genuinely useful, active job listings—including a software editor position at CNET—complete with direct links to the applications.
• Meta Muse and Gemini Spark: Both located the resume, and both fumbled the search. Instead of direct, active job listings, Muse and Spark handed back generic job-aggregator search pages or expired postings. Spark's advice was particularly phoned-in, essentially suggesting that the reviewer go search for senior tech writer jobs.

There was, however, a fascinating plot twist under the hood. When searching for the resume, both Dots and Muse pulled an older draft sitting in the Gmail inbox. Spark, thanks to Google's "Personal Intelligence" integration across the Google ecosystem, automatically reached into the reviewer's Google Drive, found the most up-to-date resume file stored there, and used that instead. It still failed to find good job links, but its contextual awareness of the user's file system was superior.

3. Video Game Wiki Scraping and Spreadsheets​

Standard AI chatbots frequently choke when asked to pull structured data from sprawling, script-heavy websites or gaming wikis. To test whether browser-equipped agents could break through that wall, PCMag asked all three to visit the official Warframe wiki, identify the 10 most recently released characters, find where their component parts drop, list the best missions to farm them, and organize the whole thing into a clean spreadsheet.

• The Result: A clean, three-way tie. Because all three agents could render and navigate the wiki pages through a browser rather than relying solely on static search snippets, Dots, Spark, and Muse all compiled the requested drop tables and formatted the spreadsheets accurately.

4. Booking a Dinner Reservation​

The fourth test asked the agents to find a well-reviewed local Asian restaurant with an open table at 6:00 p.m. that same evening.

• ChatGPT Dots: Nailed the assignment. It tracked down a well-reviewed nearby spot, found an open slot at 6:15 p.m., and explicitly noted that this was the closest available reservation to the requested 6:00 p.m. window.
• Meta Muse: Refused to pick a single winner, offering three suggestions instead—none of which actually accepted reservations. Bizarrely, Muse then prompted the user to sign in to OpenTable anyway.
• Gemini Spark: Despite Google adding OpenTable to Spark's connected apps roster in July 2026, Spark also ignored the instruction to pick a single restaurant, returning a list where two of the options didn't take reservations at all.

5. The Great Inbox Unsubscribe Purge​

Finally, the agents were pointed at a Gmail inbox and told to identify marketing clutter and unsubscribe from the senders.

• ChatGPT Dots: Turned in a genuinely impressive shift. The Dot combed through more than 15 years of promotional emails, assembled a massive spreadsheet cataloguing over 700 marketing senders, narrowed that list down to the prime candidates for removal, and fired up its virtual browser to click through third-party unsubscribe pages. While it still couldn't successfully navigate every single sender's unsubscribe flow, it did the heavy lifting.
• Meta Muse: Identified emails but failed to launch its browser to actually follow through on external unsubscribe links when required.
• Gemini Spark: Essentially handed the homework back to the user. Even though Google's own help documentation explicitly markets inbox decluttering and unsubscribing from email lists as a flagship Spark use case, in PCMag's test Spark couldn't execute the unsubscribes itself—it merely spat out a list of links so the human could click them manually.

Section Summary: Across five everyday tests, OpenAI's Dots was the only agent that reliably bridged the gap between finding information and executing the final step. Muse and Spark succeeded at read-only web scraping for the Warframe spreadsheet, but repeatedly stumbled whenever tasks required complex multi-step filtering, booking logic, or live third-party account actions.

Under the Hood on Your PC: The Hidden Friction of Browser Agents​

Why do these agents stumble so often on tasks a human can finish in three minutes? And what actually happens on your Windows or Mac desktop when you let an AI "drive" your browser?

A closer look at Google's official Gemini Apps Help documentation for Spark reveals the substantial technical and security compromises required to make consumer agents work right now.

Local Chrome vs. Cloud Browsers​

When you assign a web task to Gemini Spark on a desktop computer, it can use Chrome auto browse to hijack your local browser session, or fall back to a remote cloud browser.

Using your local Chrome instance solves the biggest headache that plagued Meta's Muse during the Instacart test: authentication. Because your local Chrome browser already holds your active session cookies and saved credentials in Google Password Manager, Spark (with your permission) can waltz right into sites you are already logged into without begging for two-factor authentication codes every five minutes.

However, Google's documentation notes a glaring catch for desktop users: your computer and Google Chrome must remain awake and running for the entire duration of a local browser task. Close your laptop lid, let Windows go to sleep, or shut down Chrome, and the local automation breaks—forcing Gemini to either failover to a remote browser (where it will immediately pause and wait for you to manually sign in to websites) or abandon the workflow. Furthermore, Google caps users at 15 concurrent running tasks, and work or school Google Workspace accounts are currently locked out of the feature entirely, restricting it to personal Google accounts.

The Privacy and Security Minefield​

Handing an AI agent the keys to an authenticated browser session—or linking it directly to your Gmail and cloud storage—creates a massive surface area for mistakes. Look closely at what happened in PCMag's testing and what the vendors themselves warn about:

1. Unintended Financial Actions: Meta's Muse casually added unrequested saffron—one of the most expensive spices on the planet by weight—straight into an Instacart order. While agents from OpenAI and Google are designed to pause and ask for human confirmation before finalizing a purchase, enviar an email, or submitting sensitive web forms, a careless user clicking "Approve" without auditing the cart will pay for the AI's hallucinations.
2. Data Leakage to Third-Party Websites: Google's support documentation for Spark includes a remarkably candid warning: when acting on your behalf in a local or remote browser, the agent may share information from your chats, Connected Apps, and Personal Intelligence directly with third-party websites. That can include your name, contact details, private files, and personal preferences.
3. Prompt Injection and Remote Data Remnants: When an agent scrapes arbitrary websites or reads incoming emails in your inbox, it is exposed to untrusted text. Google explicitly warns in its documentation that malicious sites could attempt to trick an agent into exposing insights from your Connected Apps or running malicious commands during remote code execution. To mitigate this, users have to remember to manually purge their "remote browser data" (which stores authentication cookies) and "remote code execution data" inside their Gemini Task settings after sensitive sessions.

Section Summary: Letting an AI agent operate on your behalf isn't magic; it is a messy tug-of-war between convenience and security. Cloud browsers (like Muse and Dots) get tripped up by logins and 2FA prompts unless a direct API integration exists, while local browser hijacking (like Spark in Chrome) demands that your PC stay awake and exposes your logged-in sessions to potential data oversharing.

How to Safely Test AI Agents Without Wrecking Your Digital Life​

If you are curious enough to take Muse's free tier, Gemini Spark ($20/month on AI Pro), or ChatGPT Dots ($100/month on Pro) for a spin on your PC, treat them like an overconfident intern on their first day. Here is a practical checklist to keep your accounts and wallet intact:

1. Audit Connected App Permissions First: Before firing off a prompt, open the connected apps or extensions dashboard in ChatGPT, Gemini, or Meta AI. Only connect services you actively plan to automate, and disconnect them when you are done.
2. Never Type Passwords Into a Chat Thread: As Google's own security guidelines stress, never type login credentials, two-factor codes, or credit card numbers directly into the text prompt. Always wait for the agent to pause and use its "Take control" browser mode to enter sensitive credentials directly onto the webpage yourself.
3. Stick to Read-Only Research and Structuring: As the Warframe wiki test proved, agents excel at tedious, multi-page web scraping and spreadsheet assembly. Use them to gather prices, compare specs across ten tabs, or summarize long archives before you ever trust them to click "Buy" or "Book."
4. Watch the Local Browser Side Panel: If you use Gemini Spark inside desktop Chrome, keep the Gemini side panel open so you can see the agent's planned steps in real time. Keep your finger near the Stop button in case it wanders onto the wrong site or misinterprets an instruction.
5. Purge Remote Session Cookies: If your agent uses a cloud computer or remote browser to log into your accounts, periodically go into your agent's task settings and wipe saved remote browser cookies and execution files so your authentication tokens aren't sitting indefinitely on a remote virtual machine.

The Verdict: A $100 Solution Looking for a Problem?​

There is no denying that consumer AI agents have matured over the past year. Having ChatGPT's GPT-6 Astra-powered Dots comb through 15 years of spam, build a 700-row spreadsheet of marketing offenders, and autonomously navigate external unsubscribe pages is a genuine glimpse of the future that PC power users have been promised.

Yet PCMag's hands-on gauntlet makes it painfully obvious why AI agents are still far from essential. The one agent that consistently worked—ChatGPT Dots—is locked behind a $100-per-month Pro subscription (or enterprise licensing) and still moves at a deliberate, unhurried pace while occasionally missing edge cases. Meanwhile, the more affordable $20-per-month Gemini Spark and free Meta Muse routinely choked on basic instructions, froze the Chrome browser, recommended restaurants that don't take reservations, or handed the user a list of links to click themselves.

Unless your daily workflow involves hours of repetitive, low-stakes web scraping that can be safely run in a sandboxed cloud browser, paying $1,200 a year for an AI agent to slowly order risotto ingredients or halfway clean your inbox simply doesn't pencil out. For now, a standard chatbot prompt and five minutes of your own mouse-clicking will still beat a fleet of cute AI mascots almost every time.