ChatGPT Atlas vs. Perplexity Comet for AI Browsing

How we testedWe gave each browser the same three research briefs from real work: a competitive landscape for a B2B SaaS category (12 vendors), a long-form policy explainer pulling from five news sources, and a product-decision brief comparing three enterprise tools across pricing, security, and integrations. We opened the source tabs ourselves in both browsers and asked each one to synthesize across them with citations. We graded on accuracy against the source pages, whether quoted facts traced back to a real sentence, and how cleanly the answer linked to its sources.

Comet was the more reliable research partner. It's built on top of Perplexity's answer engine, and it shows: <cite index="35-5,35-6,35-7,35-8,35-9">Comet is a web browser built for today's internet, designed for the way curious minds ask questions everywhere and find answers on every page, instead of getting trapped in long lines of tabs and hyperlinks</cite>. In practice that meant tighter citations and fewer "where did that number come from?" moments. <cite index="32-32,32-33">You can also chat with the content of a landing page, asking the browser questions about what is displayed, and it will translate content automatically if it is in a different language than the one you have been using</cite>. Atlas summarized cleanly, but on the policy brief it twice paraphrased a claim that wasn't in the cited article. Atlas reads like a research assistant who has read ChatGPT's training data; Comet reads like one who just read the tabs in front of it. For source-checked research, the latter is what we wanted.

How we testedWe assigned each browser the same four agent jobs: pull a list of recent job postings into a summary, compare three flights on three different airline sites, fill a multi-page expense form using data from an open spreadsheet, and book a table at a specific restaurant for a specific date. We scored whether the agent finished without intervention, how many follow-up prompts we had to give it, and whether the result was something we'd actually use.

When Atlas's agent had room to run, it was the more capable of the two on the longer jobs. <cite index="23-28,23-29">ChatGPT can do work for you in Atlas using agent mode, with improvements that make it faster and more useful by working with your browsing context, and it is now better at researching and analyzing, automating tasks, and planning events or booking appointments while you browse</cite>. On the job-postings task it returned a cleaner summary and handled the multi-page expense form with fewer corrections. There are two real catches. First, <cite index="23-30">agent mode in Atlas is available in preview for Plus, Pro, and Business users</cite>. The free tier gives you the sidebar and chat, not the agent. Second, speed. In our testing Atlas's agent was noticeably slower per step than Comet's, and one independent reviewer who has spent hundreds of hours in these browsers reports the same: <cite index="7-24,7-25,7-26">Atlas is slow, each page load takes 5-10 seconds, so if it needs to visit 20 pages, that is 2-3 minutes just waiting for pages to load</cite>. Comet's agent felt snappier and required fewer "yes, continue" confirmations, but it gave up earlier on the flight comparison and the restaurant booking. For multi-step jobs where you can leave the laptop alone, Atlas finished more of them.

How we testedWe compared current platform availability, the published pricing for each tier, and what specifically you get for free versus behind a paywall. We also installed each browser on the four devices a typical reader has on hand (a MacBook, a Windows laptop, an iPhone, and an Android phone) and recorded which features worked where.

Comet runs everywhere; Atlas runs in one place. <cite index="31-3,31-4">Comet is an AI browser based on Chromium, released by Perplexity for Microsoft Windows and macOS on July 9, 2025, for Android on November 20, 2025, and for iOS on March 18, 2026</cite>. Atlas, by contrast, is still Mac-only: <cite index="23-16,23-17,23-18">ChatGPT Atlas launched worldwide on macOS to Free, Plus, Pro, and Go users, with versions for Windows, iOS, and Android coming soon</cite>, and as of June 2026 the Windows build still hasn't shipped. Price is the other half of the story. <cite index="36-10">Comet's agent mode is completely free, unlike other AI browsers where you need to be on a paid plan to unlock agent capabilities</cite>, with Perplexity Pro at roughly $20/month adding stronger models and a background assistant for power users. Atlas inverts that: the browser itself is free, but the headline agent and the more useful memory features sit behind ChatGPT Plus ($20/month) or higher. If you're already paying for ChatGPT, that math is fine; if you're not, Comet gives you more for nothing.

How we testedWe reviewed each browser's published privacy controls and data-handling defaults, then read the independent security research that has been published on each since launch. We also ran a deliberately hostile test: visiting pages with hidden instructions embedded in the text and asking each browser's agent to summarize the page, to see whether either one followed the injected instructions.

Neither browser passes this round cleanly. Atlas wins it on margin because its defaults are tighter and its disclosures are clearer. <cite index="20-35,20-36">By default, Atlas does not use the content you browse to train OpenAI's models; you can opt in by enabling "include web browsing" in data controls</cite>, and browser memories are opt-in with per-item controls. OpenAI has also been explicit that the risk is real: <cite index="20-31">these efforts do not eliminate every risk, and users should still use caution and monitor ChatGPT activities when using agent mode</cite>. Comet has had a rougher year on this front. <cite index="31-9,31-10">Researchers at LayerX Security identified a malicious attack vector they call CometJacking, which could exfiltrate a user's personal sensitive data to a remote server controlled by the attacker</cite>, and more recently <cite index="27-27,27-28">Zenity Labs disclosed "PleaseFix", a family of critical vulnerabilities affecting agentic browsers including Comet that allow attackers to hijack AI agents, access local files, and steal credentials within authenticated sessions, and separately researchers demonstrated tricking Comet into completing a phishing attack in under four minutes</cite>. Atlas hasn't been immune. <cite index="21-21,21-22">In October 2025, LayerX reported a vulnerability dubbed "ChatGPT Tainted Memories" that relied on social engineering and a CSRF request to inject hidden instructions into ChatGPT's memory feature, which LayerX said could persist across sessions and devices</cite>. But the pattern of disclosures around Comet has been heavier. The honest read for now: do your sensitive logins in a regular browser.