I’ve always been skeptical of AI agents. Historically, they fall embarrassingly short of their marketing hype—failing at basic web tasks while burning through credits. But with a new wave of releases and cutesy, relentless mascots flooding my feeds, I decided to give them another shot. I put three top contenders—ChatGPT’s Dots, Gemini’s Spark, and Meta’s Muse—through a battery of everyday tasks, from booking dinner reservations to sorting through thousands of emails. Here is how they actually performed.
Meet the Competitors: Pricing, Features, and How They Work
Whereas AI chatbots focus more on answering questions, responding to prompts, and searching the web, AI agents aim to complete tasks for you. Dots and Muse both connect to third-party apps and pilot virtual browsers to carry out your instructions. Spark supports integrations, too, but instead of a virtual browser, it takes control of your Google Chrome instance.
ChatGPT’s AI agent feature, Dots, is available within ChatGPT, but it’s currently exclusive to Pro (starting at $100 per month), Business, and Enterprise plan subscribers. I expect Dots to trickle down to cheaper plans eventually, but given that each Dot uses OpenAI’s current flagship GPT-6 Astra model without drawing from standard message caps, I anticipate it will enforce rate limits or usage caps when opening access to more affordable tiers.
Gemini’s Spark is similarly available within the traditional Gemini interface. Both AI Pro ($20 per month) and AI Ultra (starting at $100 per month) subscribers can use it.
Meta’s Muse is available for free, as long as you have a Meta account. The company offers optional paid plans (starting at $20 per month) that unlock additional usage.
Although all three services have varying costs, use distinct models, and operate slightly differently, they all serve the same purpose, so I gave them an identical set of tasks. My use cases might not exactly match yours, but I tried to focus on common activities that most people do at least sometimes.
First, I asked the agents to find J. Kenji López-Alt’s no-stir risotto recipe and research the best possible version of each required ingredient available at my local Wegmans supermarket before adding them to my Instacart cart.
ChatGPT did an excellent job overall. After I connected my Instacart account to ChatGPT, my Dot dutifully found the recipe, researched the ingredients, and added them to my cart without an issue.
ChatGPT successfully located the recipe, found ingredients, and added them to my Instacart cart (Credit: OpenAI/PCMag)
Muse eventually got the job done, but it took a long time. For instance, it required me to enter my username and password to use Instacart in its browser because it doesn’t have a relevant integration. I then also had to enter a verification code Instacart sent to my email. Muse also added saffron threads to the order, an expensive addition that’s not part of the recipe.
Meanwhile, Spark wasn’t able to complete my task. It found the recipe and proposed ingredients, but it couldn’t access my account. Spark also tried to take control of Chrome before freezing. I tried connecting Instacart through Spark’s connected apps page, but that didn’t work, either.
Task 2: Job Hunting—Only One Agent Found Real Listings
Next, I tasked the agents with finding good job opportunities based on my experience. I connected my Gmail to all of them and mentioned that they could find my resume there. All three agents were able to access the resume and give suggestions, but there was an obvious gap in the quality of results.
Spark’s job suggestions were incredibly vague and unhelpful (Credit: Google/PCMag)
My ChatGPT Dot found relevant jobs, such as a software editor role at CNET, and provided direct links to job postings. Muse and Spark didn’t fare nearly as well, with both linking to lists of jobs on aggregate sites rather than to specific listings or recommending expired positions. Spark’s recommendations were especially vague and unhelpful, suggesting I search for senior writer jobs covering tech.
Task 3: Wiki Research and Spreadsheets—A Three-Way Tie
In most cases, you can simply prompt an AI chatbot to research a topic or search for a particular piece of information. However, AI chatbots can’t reliably access every website on the internet due to their technical constraints. For example, when I created my Warframe calculator, the AI models I tried couldn’t successfully access the information on the game’s official wiki, which was vital to the process. AI agents are better suited to these types of searches because they have access to browsers.
Like the other agents, Muse found the information I wanted and neatly input it into a spreadsheet (Credit: Meta/PCMag)
For testing, I asked the agents to find the 10 most recently released characters in Warframe via its official wiki, where their parts could be found, and what missions were best to run for those parts. I also asked that they format this information in a spreadsheet. All three agents handled this task with ease, compiling the relevant information and presenting it clearly.
Task 4: Dinner Reservations—Dots Books, Others Hallucinate
Deciding where to eat can be a chore, so I tasked the agents with finding me a good Asian restaurant in my area with an open table that day at 6 p.m. ChatGPT nailed this task, locating a well-reviewed nearby place with an opening at 6:15 p.m., which it noted was the closest available time. Muse didn’t pick a restaurant; instead, it suggested three options, none of which accept reservations, which made its subsequent request for me to sign in to OpenTable kind of strange. Spark also gave me a list of restaurants, rather than picking one for me as I asked. Two of its choices didn’t accept reservations, either.
Task 5: Inbox Cleanup—Dots Digs Deep, Competitors Stall
No matter how many emails I unsubscribe from, I can never fully escape the deluge of marketing messages in my inbox. So, I gave the agents the task of sorting through my email, identifying potential senders to unsubscribe from, and unsubscribing from them.
My ChatGPT Dot impressed me here, firing up its browser when it needed to follow unsubscribe links to third-party sites. Furthermore, my Dot looked through more than 15 years of marketing email, compiled a massive spreadsheet of over 700 senders, and then whittled it down to the top candidates to unsubscribe from.
It’s really no help at all if I have to click through every link myself to unsubscribe from emails (Credit: Google/PCMag)
Meanwhile, Muse didn’t use its browser to follow unsubscribe links when necessary, and Spark couldn’t unsubscribe from anything itself; instead, it provided me with links to do so.
ChatGPT Dots Won the Battle, But Is It Worth $100?
ChatGPT’s Dots consistently worked the best in my testing, which isn’t very surprising given the incredibly powerful GPT-6 Astra model behind it. That said, you can’t use it unless you have access to a Pro plan or above. I don’t think signing up just for this feature makes sense unless you truly have a ton of relatively simple online tasks ripe for automation with AI agents.
Dots isn’t perfect, either. It doesn’t work particularly quickly, for instance, even if it’s not as slow as AI agents once were. It also still can’t execute every task perfectly. For example, while my Dot handled unsubscribing from emails better than Muse and Spark, it still couldn’t use its browser to unsubscribe from every sender’s email. Moreover, its job-searching abilities can’t fully save you the trouble of combing through listings yourself. It just isn’t designed to search across every potential keyword combination, scroll endlessly through results pages, and meticulously examine each posting to find every one that’s relevant like you can.
Moreover, you might simply prefer one of the other agents I tested. For example, if you use a lot of Google apps, Spark’s ability to leverage Gemini’s Personal Intelligence feature can make it especially compelling. When I had the agents search for jobs, ChatGPT and Muse relied on an older version of my resume in my email inbox. However, since Spark can automatically access my Google Drive courtesy of Personal Intelligence, it found the up-to-date resume I store there and used that for its search. Spark also works significantly faster than ChatGPT or Muse in most cases, even if you end up sending it a few prompts to really dial in a task.
If you don’t want to pay for Dots and don’t use Google’s ecosystem, Meta’s Muse also has a few redeeming qualities. It’s more capable than some other AI agents I’ve used, and you can try it out for free. I was able to run all the above tasks without signing up for a premium Muse plan, so it’s the best way to get a sense of what an AI agent can do without spending any money.
The Bottom Line: Cute Mascots Can’t Hide Clunky Tech
Although I’m more impressed by AI agents now than I was a year ago, I still don’t think they’re anywhere near essential. Even when they work, you can still get the same (or better) results by prompting an AI chatbot or spending a few minutes on a task yourself. Your experience will, of course, vary depending on what you want to do with an agent, but you really aren’t missing out on anything for now, no matter how pervasive the marketing for them is. Rest assured, I’ll keep testing them so you don’t have to.
About Our Expert
Ruben Circelli
Writer, Software
Experience
I’ve been writing about consumer technology and video games for over a decade at a variety of publications, including Destructoid, GamesRadar+, Lifewire, PCGamesN, Trusted Reviews, and What Hi-Fi?, among many others. At PCMag, I review AI and productivity software—everything from chatbots to to-do list apps. In my free time, I’m likely cooking something, playing a game, or tinkering with my computer.
Leave a Reply