I Stress-Tested the New OpenAI Atlas Browser

In partnership with

DEEP DIVE
I Stress-Tested the New OpenAI Atlas Browser

I called it a few months ago here.

AI browsers are the next step for actually using AI in your day-to-day work.

OpenAI just shipped their version: Atlas, months after Perplexity’s Comet release.

Recap

Here’s what this thing is, in plain terms.

Using ChatGPT in a browser tab (like Chrome) has a hard limitation:

It can’t see what you’re doing or take action on the actual page.

So OpenAI (like other AI labs) built its own browser around their LLM.

• You get your chat history on one side, and the page you’re currently on on the other.

• The agent can take actions like clicking, typing, navigating, and logging in (with your permission.)

• It still leans on Google for search results (which is kind of ironic, but whatever).

• There’s also a “browser memory” feature, which works like an AI-powered search history. It remembers what you’ve done and lets you pick things back up later.

So that’s the setup.

Cool, but what can I actually do with it?

I ran a bunch of tests for real tasks I already do in my business.

Here’s how it performed.

Test #1: Buy something

Goal: Recommend and buy beginner climbing shoes for me.

Result: Took about 5 minutes to figure out the site. It eventually added a pair of beginner shoes to the cart. It struggled to understand some elements, but didn’t completely fail.

✅ It did the job, but slowly. About the same time as a VERY slow human.

Test #2: Find logos

Goal: Find transparent PNG versions of logos from a list I gave it.

Result: Fast and accurate. Found good quality logo files and gave me direct links to download them.

✅ This was solid. Took about the same time it would’ve taken me to do manually.

Test #3: Apply to a freelance job

Goal: Go to my Contra profile and apply to the most relevant recent freelance job.

Result: It needed to log in. Took a while. Didn’t do exactly what I asked, but it did navigate the site and draft an application.

⚠️ It wasn’t quite what I wanted, but it understood the intent.

Test #4: Website management

Goal: Identify low-performing articles from a traffic report and unpublish them in WordPress.

Result: Took about 7 minutes. It didn’t want to do it at first (“I can’t help with that”), but after some nudging, it eventually followed through and completed the task.

✅ Technically successful, but painfully slow. I could’ve done it in 30 seconds.

Test #5: Data and research

Goal: Go through the last 20 newsletters in my inbox, find the sponsors, and add them to a Google Sheet for tracking.

Result: Failed. The browser agent couldn’t click the email links properly due to spam protection and security prompts. It got stuck.

❌ Unsuccessful.

TL;DR – How it did overall

The agent can technically do a lot of things. But right now, it sits in an awkward spot:

• For small tasks, it works, but usually takes 3x longer than if you did it yourself.

• For big tasks, the kind that would actually save you time, it either breaks, gives up, or needs too much hand-holding

This wasn’t a full-blown test suite. These were just my own tasks. But if you’re curious, try running it through some of the boring stuff you do every week.

That’s where you’ll really see whether it’s ready, or just pretending to be helpful.